Though it could expand his outlet’s reach — especially, in his words, as email becomes increasingly important “as a distribution channel with social media declining as a traffic source” — Brodbeck didn’t think creating an additional newsletter was an optimal use of reporter time in the zero-sum, resource-strapped reality of running a hyperlocal news outlet.

“As much as I would love to have a 25-person newsroom covering Northern Virginia, the reality is that we can only sustainably afford an editorial team of eight across our three sites: two reporters/editors per site, a staff [photographer], and an editor,” he said. In short, tapping a reporter to write a morning newsletter would limit ARLnow’s reporting bandwidth.

But with the exponential improvement of AI tools like GPT-4, Brodbeck saw an opportunity to have it both ways: He could generate a whole new newsletter without cutting into journalists’ reporting time. So last month, he began experimenting with a completely automated weekday morning newsletter comprising an AI-written introduction and AI summaries of human-written stories. Using tools like Zapier, Airtable, and RSS, ARLnow can create and send the newsletter without any human intervention.

Since releasing the handbook, Amditis has heard that many publishers and reporters “seem to really appreciate the possibility and potential of using automation for routine tasks,” he told me in an email. Like Brodbeck and others, he believes “AI can save time, help small newsrooms scale up their operations, and even create personalized content for their readers and listeners,” though he raised the widely held concern about “the potential loss of that unique human touch,” not to mention the questions of accuracy, reliability and a hornets’ nest of ethical concerns.

Even when instructing AI to summarize content, Amditis described similar challenges to those Brodbeck has encountered. There’s “a tendency for the summaries and bullet points to sound repetitive if you don’t create variables in your prompts that allow you to adjust the tone/style of the responses based on the type of content you’re feeding to the bot,” he said.

But “the most frustrating part of the work I’ve been doing with publishers of all sizes over the last few months is the nearly ubiquitous assumption about using AI for journalism (newsletters or otherwise) is that we’re out here just asking the bots to write original content from scratch — which is by far one of the least useful applications, in my opinion,” Amditis added.

Brodbeck agrees. “AI is “not a replacement for original local reporting,” he said. “It’s a way to take what has already been reported and repackage it so as to reach more readers.”

Specifically, it looks a lot like the SEO-driven, human-written, meh content that you’ll find all over the rest of the internet. As Noor Al-Sibai and Jon Christian reported for Futurism on Thursday:

The 40 or so articles, all of which appear to be SEO-driven travel guides, are comically bland and similar to one another. Check out these almost-copied lines:

• “Now, I know what you’re thinking – ‘Cape May? What is that, some kind of mayonnaise brand?'” in an article about Cape May, in New Jersey.
• “Now I know what you’re thinking – ‘but Caribbean destinations are all just crowded resorts, right?'” in an article about St Maarten, in the Caribbean.
• “Now, I know what you’re thinking. Puerto Rico? Isn’t that where all the cruise ships go?” in an article about San Juan, in Puerto Rico.
• “Now, I know what you’re thinking- bigger isn’t always better,” in an article about Providence, in Rhode Island.
• “Now, I know what you’re probably thinking. Nepal? The Himalayas? Haven’t we all heard of that already?” in an article about Khumbu, in Nepal.
• “Now, I know what you’re probably thinking. “Brewster? Never heard of it,” in an article about Brewster, in Massachusetts.
• “I know what you’re thinking: isn’t Stockholm that freezing, gloomy city up in the north that nobody cares about?” in an article about Stockholm, in Sweden.

That’s not the bot’s only lazy trope. On review, almost everything the bot has published contains at least one line about a “hidden gem.”

You can see all the articles (THAT WE KNOW OF) here; they’re bylined “As Told to Buzzy,” a winking bow-tied robot with the bio “Articles written with the help of Buzzy the Robot (aka our Creative AI Assistant) but powered by human ideas.” Forty-four of them were published this month.

The articles are pretty bad.

this stuff is GOLD, what is everyone worried about!!! pic.twitter.com/3JkqkoEFQB

— Kathryn VanArendonk (@kvanaren) March 30, 2023

BuzzFeed is publishing AI-generated travel articles.

On the one hand, they're terrible. On the other, they're just reflecting back the fact that most travel writing online is truly awful.https://t.co/oUA0qX5uXu pic.twitter.com/fvCIDvDAN8

— Max Tani (@maxwelltani) March 30, 2023

credit where credit's due, this is an astonishing lede pic.twitter.com/JfTfumLSQy

— Kathryn VanArendonk (@kvanaren) March 30, 2023

Now, I know what you’re thinking: The concern isn’t that these specific articles are going to win any awards, it’s that AI’s future potential is so great that this is simply the tip of the iceberg and human writers will soon be replaced by a sentient, independent Buzzy.

Maybe. But another way to look at it is that a lot of the human-written content that Buzzy’s articles are competing with are also pretty bad — “essentially human-made AI”:

it was bc there was the need for more and constant content stream. so writers would scour the internet for what's been written before and essentially just rewrite it, maybe give it a new placement. but it was all the same, very random pool. same terminology, etc.

— cale g weissman (@caleweissman) March 30, 2023

anyway, now AI is actually doing this and it's worse in a different but funnier way: https://t.co/3M22BYhsnA

— cale g weissman (@caleweissman) March 30, 2023

A BuzzFeed spokesperson told Futurism that the company is using Buzzy + a human editor “to unlock the creative potential of UGC so we can broaden the range of ideas and perspectives that we publish,” with people picking the topics (in this case, specific cities) and Buzzy doing the, um, generating.

It’s not that different from a freelance assignment I did in my twenties: A human editor assigned me to write some articles about the promise of 5G — a topic about which I knew nothing — and I googled 5G, read other content mill-ish articles about it, and compiled them into my “own” article. The content I created wasn’t really meant to be read by humans who actually needed to know anything about 5G, in the same way that anybody who is planning a trip to Morocco probably shouldn’t get their recommendations from Buzzy. (Its recommendations are: Go to Marrakesh, the mountains, and the desert. It’s far away. Bye!)

In the 5G case, I was basically Buzzy, except I was getting paid. Buzzy works for free. You know, for now.

"They're taking the jobs of severely underpaid offshore 'content writers' for chum-bucket sites that no one reads except for ad-fraud bots!" If they can fake the money, it's a perfect closed circle.

— Steven Scott (@scottstev) March 30, 2023

Very formulaic picture-driven travel listicles. But this experiment shows that the writing is on the wall for anyone currently making a living from generating that type of content. https://t.co/uWAxfxJCyi

— Peter Aldhous (@paldhous) March 30, 2023

1. Technology XYZ is arriving. It will be incredible for everyone. It is basically inevitable.

2. The only thing that can stop it is regulators and/or incumbent industries. If they are so foolish as to stand in its way, then we won’t be rewarded with the glorious future that I am promising.

We can think of this rhetorical move as a Reverse Scooby-Doo. It’s as though Silicon Valley has assumed the role of a Scooby-Doo villain — but decided in this case that he’s actually the hero. (“We would’ve gotten away with it, too, if it wasn’t for those meddling regulators!”)

The critical point is that their faith in the promise of the technology is balanced against a revulsion towards existing institutions. (The future is bright! Unless they make it dim.) If the future doesn’t turn out as predicted, those meddlers are to blame. It builds a safety valve into their model of the future, rendering all predictions unfalsifiable.

This trope has been around for a long time. I teach it in my history of the digital future class, with examples from the ’90s, ’00s, and ’10s. It’s still with us in the present day. And lately it has gotten intense.

Take a look at this tweet from Balaji Srinivasan. Click and you can read the whole thread. Srinivasan is a popular venture capitalist with a significant following among the tech class. He’s the author of The Network State, a book that you probably shouldn’t read. (The Network State isn’t a good book, but it is a provocative book, in much the same way that Elon Musk buying Twitter for $44 billion wasn’t a good investment, but it sure does make you think.) He writes a lot about the inevitability of crypto, AI, and everything else in his investment portfolio. (Balaji was a big fan of Clubhouse. Remember Clubhouse?)

AI means a brilliant doctor on your phone. Who can diagnose you instantly, for free, privately, using only your locally stored medical records.

Do you think the doctors will be happy about that? Or the lawyers? The artists? The others that AI disrupts?

They’ll fight it. Hard.

— Balaji (@balajis) February 10, 2023

AI directly threatens the income streams of doctors, lawyers, journalists, artists, professors, teachers.

That happens to be the Democrat base!

So they’ll lash out. Hard. AI safety for them means job security. Everything is on the table, from lawsuits to laws.

— Balaji (@balajis) February 10, 2023

Let’s break down what he’s doing in this tweet thread. He’s stringing together two empirical claims to establish the trajectory of an ideological narrative.

Claim 1: AI means a brilliant doctor on your phone, for free.

Claim 2: AI directly threatens the income streams of doctors, lawyers, journalists, etc. Their industries will resist attempts at AI-based disruption.

The ideological narrative: These entrenched interests are going to try to short-circuit the awesome potential of AI. Democrats in government will go along with them. We ought to oppose them today, and blame them for any shortcomings tomorrow.

(That right there? That’s a Reverse Scooby-Doo, folks.)

The first claim is not even a little bit true. AI is not, at present, a “brilliant doctor on your phone, for free.” It is nowhere close to that. There are few stupider use cases for the current crop of generative AI tools than asking them to diagnose non-obvious, potentially-critical medical symptoms. Recent attempts to deploy machine learning to aid COVID response went disastrously awry. There is an established track record here. It’s terrible. AI is optimistically decades away from being suitable for such a task. It might never be an appropriate use case.

Balaji is simply projecting, insisting that in the future, AI companies will surely solve those problems. This is a type of magical thinking. And like all real magic, what they are actually attempting is an elaborate misdirection.

Consider: If AI is ever going to become your instant free doctor, the companies developing these tools are going to require a truly massive dataset. They’ll need limitless access to everyone’s medical records.

The implicit plan Srinivasan is pushing looks something like this:

Step 1: Give up any semblance of medical privacy.
Step 2: Trust startups not to do anything shady with it.
Step 3: TKTK, something about Moore’s Law and scientific breakthroughs. We’ll work all that out later.
Step 4: Profit!

Fake-it-till-you-make-it hasn’t gone great for medical tech startups. The last big one to try was Theranos, and the executives of that company (Elizabeth Holmes and Sunny Balwani) are now serving 11 and 13 years in prison, respectively. So Balaji’s imagined future only has a chance if he can divert attention away from the pragmatic details.

Now there’s actually a version of his second empirical claim that I agree with. (Hell, I made a similar argument a couple months ago.) I expect well-credentialed industries will be much less impacted by developments in generative AI than industries that are mostly made up of freelancers. Lawyers will be fine; digital artists are going to face a world of hurt.

But this isn’t because “they’re the Democrat base.” It’s because well-credentialed industries are positioned to represent and protect their own interests.

Lawyers and doctors are the two obvious examples here. An AI might be able to correctly diagnose your symptoms. But it cannot order medical scans or prescription drugs. Insurers will not reimburse medical procedures on the basis of “ChatGPT said so.” An AI could also write a legal contract for you. Hell, you could probably track down boilerplate legal contract language using an old-fashioned Google search too. But that will work right up until the moment when you need to enforce the contract. That’s when you run the risk of learning you missed a critical loophole that a savvy lawyer who specializes in the actual field would know about.

When billionaire tech entrepreneurs like Balaji insist that AI will replace lawyers, let’s keep in mind what they actually mean is AI will replace other people’s lawyers. (Just like Elon Musk doesn’t intend to live on Mars. He wants other people to colonize Mars for him.)

It brings me back to William Gibson’s famous dictum: “The future is already here—it’s just not evenly distributed.” I’ve written about this previously, but what has always stood out to me is that the future never becomes evenly distributed. Balaji and Marc Andreessen and Sam Altman aren’t living in or constructing a future that everyone else will eventually get to equally partake in. The uneven distribution is a persistent feature of the landscape, one that helps them to wield power and extract audacious rents.

Srinivasan isn’t so much making empirical claims here as he is telling a morally-charged story: Pledge your allegiance to the ideology of Silicon Valley. Demonstrate faith in the Church of Moore’s Law. All will be provided, so long as the critics and the incumbent industries and the regulators stay out of the way. Faith in technological acceleration can never fail, it can only be failed.

And Balaji is hardly alone here. This type of storytelling has a strong pedigree in the archives of digital futures’ past. Tech ideologues have been weaving similar tales for decades.

In 1997, Wired magazine published a bizarre tech-futurist manifesto of sorts, “Push!” The magazine’s editorial team declared that the World Wide Web was about to end. It would be replaced, inevitably, by “push” media — companies like BackWeb and PointCast that pushed news alerts to your desktop computer and would one day reach you on every surface of your home. They envisioned “technology that, say, follows you into the next taxi you ride, gently prodding you to visit the local aquarium, all the while keeping you up-to-date on your favorite basketball team’s game in progress.”

The more closely you read “Push!” the less sense the argument makes. At one point they argue that Wired’s old-fashioned magazine is both pull-media and push-media. Never once do they consider whether email might already be a well-established form of push media. The whole thing is kind of a mystery.

But what they lacked in clarity they made up for in certainty. The authors declare that the oncoming Push! future is inevitable, because “Increasingly fat data pipes and increasingly big disposable displays render more of the world habitable for media” and “Advertisers and content sellers are very willing to underwrite this.” The web is surely dead, in other words, because Wired’s editors have seen a demo, they have a sense of some tech trends, and they are confident advertisers will foot the bill.

But then, they include this caveat: “One large uncertainty remains…If governments should be so stupid as to regulate the new networked push media as they have the existing push media, the expansion of media habitat could falter.”

(To summarize: Push! was arriving. It would be incredible for everyone. It was basically inevitable. That is, unless regulators started meddling. In that case, our glorious technological future could be denied.)

At no point did they consider that the technologies they were breathlessly hyping actually sound godawful. Advertising that follows you around a city, that nudges you to visit the aquarium even when you get in a taxi? Big ad-supported disposable displays that you can never turn off or outrun? That sounds…like something that we’d probably want regulators to curtail.

In a 2019 Wired cover story, “Welcome to Mirrorworld,” Kevin Kelly offered a surprisingly direct articulation of this perspective. It came in an essay declaring that augmented reality would soon arrive. It would be incredible for everyone. It was, basically, inevitable.

Let’s set aside whether AR has much of a future, and what that future will look like. My current answers are “maybe” and “it depends on a lot of factors that are still very unclear.” I plan to write more on the topic once there is more substance to write about. The critical passage appears late in the piece, where he articulates his ideological position on technology and regulation (emphasis added):

Some people get very upset with the idea that new technologies will create new harms and that we willingly surrender ourselves to these risks when we could adopt the precautionary principle: Don’t permit the new unless it is proven safe. But that principle is unworkable, because the old technologies we are in the process of replacing are even less safe. More than 1 million humans die on the roads each year, but we clamp down on robot drivers when they kill one person. We freak out over the unsavory influence of social media on our politics, while TV’s partisan influence on elections is far, far greater than Facebook’s. The mirrorworld will certainly be subject to this double standard of stricter norms.

As an empirical matter, Kelly’s “Mirrorworld” (a 1-to-1 digital twin of the entire planet and everything inhabiting it) is still a long way off. Like Srinivasan, what Kelly is doing in the piece is projecting — demonstrating faith that the accelerating pace of technological change means we are on the path he envisions.

What Kelly’s writing gives us is a richer taste of the ideological project these tech thinkers are collectively engaged in: Abandon the precautionary principle! Don’t apply the same old rules and regulations to startups and venture capitalists. Existing society has so many shortcomings. The future that technologists are creating will be better for everyone, if we just trust them and stay out of the way!

It’s a Reverse Scooby-Doo narrative. And, viewed in retrospect, it becomes easy to pick out the problems with this approach. Have faith in the inevitability of Push!? Of Mirrorworld? Of autonomous vehicles? Of crypto, or web3, or any of the other flights of fancy that the techno-rich have decided to include in their investment portfolio? Push! didn’t flop because of excessive regulation. The problem with autonomous vehicles is that they don’t work. Trust in crypto’s speculative bonanza turned out to be misplaced for exactly the reasons critics suggested.

My main hope from the years of “techlash” tech coverage is that we collectively might start to take the power of these tech companies seriously and stop treating them like a bunch of scrappy inventors, toiling away at their visions of the future they might one day build. Silicon Valley in the ’90s was not the power center that it is today. The largest, most profitable, most powerful companies in the world ought to be judged based on how they are impacting the present, not based on their pitch decks for what the future might someday look like.

What I like about the study of digital futures’ past is the sense of perspective it provides. There’s something almost endearing in seeing the old claims that “the technological future is inevitable, so long as those meddling regulators don’t get in the way!” — applied to technologies that had so very many fundamental flaws. Those were simpler times, offering object lessons that we might learn from today.

It’s much less endearing coming from the present-day tech billionaire class. Balaji Srinivasan either doesn’t understand the existing limits of AI or doesn’t care about the existing limits of AI. He’s rehashing an old set of rhetorical tropes that place Silicon Valley’s inventors, engineers, and investors as the motive force of history, and regards all existing social, economic, and political institutions as interfering villains or obstacles to be overcome. And he’s doing this as part of a political project to stymie regulators and public institutions so the tech sector can get back into the habit of moving fast and breaking things. (It’s 2023. They have broken enough already.)

The thing to keep in mind when you hear Balaji and his peers declaring some version of “the technological future is bright and inevitable…so long as those meddling public institutions don’t get in the way,” is that this is just a Reverse Scooby-Doo. That line of thinking originates from the villain, and for good reason. The people who say such things are ultimately up to no good.

In contrast, the public release of ChatGPT on November 30, 2022, was very much a watershed moment, and it has many of us asking (before the fact, this time!) how our jobs and industry could be changed with the advent of the natural language functions of artificial intelligence.

Not surprisingly, the huffing and puffing over the past three months about ChatGPT and other similar AI-powered platforms run the gamut from a rush to mock its shortcomings to doomsday warnings that the machines will take our jobs. What’s clear is that all the predictions and public experiments are provisional, as the technology is rapidly evolving and so many applications are still to be discovered.

My own particular contribution in these early days comes instead from looking backward. By no design of my own, I’ve found myself swimming in the AI waters for the past decade in my job as co-founder and editor of the online magazine Worldcrunch, which publishes English editions of top foreign-language journalism.

Even before we launched in 2011, we were regularly being questioned about AI, though nobody was yet calling it that. Instead, would-be investors and partners, editorial and tech colleagues all wanted to know: “…and what about Google Translate?”

Some saw burgeoning machine translation as a threat to our business model. Others saw it as an inevitable downward pull on the quality of our editorial product. My reflex back then was to dismiss the machine, pointing out some of its more outrageous errors and assuring that it could never be good enough to take the place of our 100%- human translation.

Still, in Worldcrunch’s first years, working largely with freelance translators, I would find myself occasionally “catching” those who I suspected of using the machine. Sometimes it was incoherent copy, sometimes it was blatant errors (names, for example, get translated as words: Marine the Pen, anyone?) And sometimes I would get suspicious simply because the copy came back so quickly.

The truth is some of our best translators were also already using the machine…and there was nothing to “catch.” It was a tool, and professionals knew how to use it to help them be as good (and, yes, as fast) as possible. Those who abused or misused it were rather quickly spotted, and didn’t last long working with us.

Irene Caselli, a veteran journalist and editor on our team who speaks five languages, has been leaning on machine translation tools for years — even before they were really any good. “When it first came out, I would sometimes use Google Translate just to have many of the basic words and structure laid out in the new language on the page,” she recalled. “But I would have to check word by word.”

In the years since, while the machine has gotten exponentially better (today Irene prefers DeepL), she is quite conscious about when and how to use it. The output is generally stronger on more straightforward political and informational journalism, and weaker on writerly work and stories that “change register” within the same piece. In the languages where pronouns are often omitted (Italian, Spanish, French, etc.), translation programs still tend to use the default “he,” though Caselli says they are improving and can sometimes figure out that it’s “she” based on context.

So while 10 years ago she used the technology “to be able to type less,” Caselli says now, “I can sometimes use it as a bonafide first draft where I can just go through and edit the copy, spotting any mistakes along the way.”

When Le Monde launched an English edition last year, it did so with the systematic integration of an AI-powered translation platform into the editorial process, with human oversight.

Here at Worldcrunch, one application that’s been revealed over time has been automated translation’s role in our work as editors. In many (though not all) of the languages we translate from, the machine has gotten good enough that I can refer to it when trying to rephrase a clunky sentence in a translated piece from a language I don’t know. Perhaps even more useful has been the ability to browse through entire newspapers in a variety of languages and feel confident assigning stories to translators who don’t necessarily know the kind of pieces we’re looking for. That can save a ton of time and mental energy.

Even with the advances of machine translation, plenty of translators — of all ages — still choose not to use it. And I’ll always prefer to get story pitches from seasoned journalists who speak the languages. I imagine that AI natural language tools will largely be used for purposes of speed and shortcuts, and that will differ down to the level of individual affinities and working habits.

With that said, the potential changes we’ve begun to imagine with ChatGPT for the production of news and journalism — and all across the creative industries — go beyond speed. It’s fundamentally different from the various digital bells and whistles that have been thrust upon us over the past decade or two.

The first round of public experiments have been interesting to watch. There are initial ethical questions for news companies vis-à-vis readers. Medium has established a first policy on “transparency, disclosure, and publication-level guidelines” for the use of AI language tools.

But ultimately, if we stay on the current trajectory, it’s utterly plausible that AI language tools will begin to blend into our daily workflows, similar to how Google and Google Translate have. That’s a very big deal.

These advanced automated language models get at the very essence of what we do — or at least half of it: the writing, synthesizing information, crafting stories that has always made us muttering hacks feel, well, human.

The Google-Facebook era put our earning power on the line. This runs deeper. It’s an ego thing. Will we be reduced to the machine’s fact-checker?

Yet there’s the other half of what news and journalism is about, which makes us feel human in another way — and that stands beyond the reach of the databases and algorithms. We are also doing our job (and feeling alive) when we find or figure something out first. Our digital world — and creativity itself — can be so derivative that we can forget that we’re here because every day new stuff happens. We see things, make connections, and occasionally, according to the famous dictum, do the only “real” journalism: Publish what someone else doesn’t want published.

To further soothe our fragile egos, we can borrow from another old industry dictum: If journalism is the first draft of history, the machine goes nowhere without us.

Here are some thoughts, culled from Worldcrunch’s experience with machine translation, that may be applicable to using AI natural language tools.

— Maximizing the utility of automation requires human reasoning/thinking/creativity before feeding it to the machine, and human oversight (which may include more reasoning/thinking/creativity) after it comes out the other side.

— Don’t manage down: Online tools are best left in the hands of individuals.

— Editors will have to rely on the same “red flag” instincts that catch sloppiness, laziness, plagiarism, etc. (though some tools to help keep up in the early days would be nice!)

— Include regular training and an open conversation about new ways to use it, and possible pitfalls (we haven’t done enough of this with other internet tools).

— If the tools are powerful and reliable enough to integrate into the editorial process, publicly labeling work as “Produced with AI,” etc., will ultimately be pointless.

— Factor in exponential improvement in quality and precision.

— Factor in that human oversight will always be necessary.

— Speed matters.

— Quality matters more.

—Accuracy matters most.

When not covering AI for the FT, Murgia is finishing her first book, Code-Dependent, out in February 2024. We caught up via email. Our back-and-forth has been lightly edited for clarity and that British proclivity for the letter “zed.”

The Wall Street Journal reported Thursday on a staff memo at BuzzFeed that laid out plans for the digital media company to use OpenAI — creator of ChatGPT — to help write quizzes and other content. In the memo, BuzzFeed CEO Jonah Peretti wrote AI will play a role in both editorial and business operations at BuzzFeed within the next year.

“For example, a quiz to create a personal romantic comedy movie pitch might ask questions like, ‘Pick a trope for your rom-com,’ and ‘Tell us an endearing flaw you have,'” the Journal’s Alexandra Bruell reported. “The quiz would produce a unique, shareable write-up based on the individual’s responses, BuzzFeed said.”

But, hey! Humans will still provide “cultural currency” and “inspired prompts,” according to Peretti’s memo.

“If the past 15 years of the internet have been defined by algorithmic feeds that curate and recommend content, the next 15 years will be defined by AI and data helping create, personalize, and animate the content itself,” Peretti wrote.

Maybe it’s because the announcement comes as several news organizations announced layoffs and other cuts, but many found the update grim.

a sentence to put on journalism’s tombstone if I’ve ever seen one: “BuzzFeed remains focused on human-generated journalism in its newsroom, a spokeswoman said Thursday.” https://t.co/QWYpwzHf9v

— Samit Sarkar (@SamitSarkar) January 26, 2023

Attention Twitter and BuzzFeed corporate: I am actually an AI chatbot. I produce quality, human-adjacent content. I pass the Turing test. I very rarely go to the doctor. I plan vacation days well in advance! Union? What union?! [muffled screaming]https://t.co/Yvj2SKk6C6

— Matt Shuham (@mattshuham) January 26, 2023

Somebody’s gonna end up with the job of editing this garbage. Will have to be the most soulkilling new job in media since whoever signed up to revive zombie Deadspin. https://t.co/Ul8PHm68E0

— Kevin Collier (@kevincollier) January 26, 2023

The stock market on the other hand? $BZFD ultimately jumped 120% on the news that the company plans to use AI to generate content, its biggest gain since going public in December 2021.

Two recent WSJ stories about BuzzFeed’s business come ahead of this stock jump, by @jtoonkel + @alexbruell

Last night, ~ $10m deal with Meta for creator content: https://t.co/erMv5OyhsN

Today, using ChatGPT to assist with content creation: https://t.co/aTUB9YxMsc$BZFD

— Kerry Flynn (@kerrymflynn) January 26, 2023

The AI-powered chatbot that can generate humanlike text on most prompts was released in late November 2022 and had a million users within a week. But we’re still learning about how it works — and how it came to be. (Time magazine, as one example, recently revealed OpenAI paid workers in Kenya less than $2 an hour to wade through some of the darkest parts of the internet.)

In one recent case of AI-powered articles gone wrong, the outlet CNET had to issue “substantial” corrections, respond to accusations of plagiarism, and ultimately hit pause on their whole AI experiment earlier this month. BuzzFeed must be hoping that using similar technology for quizzes will be less fraught.

Random quizzes and lists is exactly the type of content you should use AI to create. This is fine. https://t.co/3pOwI80U68

— Ross Maghielse (@Maghielse) January 26, 2023

Some predicted the tech could be “a game-changer” for journalism, particularly resource-strapped local newsrooms. Others cautioned that producing convincing disinformation just got a lot cheaper and faster, raised ethical questions the news industry is only beginning to grapple with, and predicted that AI-written content will soon flood the internet.

A prediction by Gannett’s Eric Ulken read, in part, “I don’t imagine we’ll see GPT-3–produced copy in the pages of The New York Times in 2023, but it’s likely we’ll turn to machines for some previously unthinkable creative tasks.”

I didn’t realize how close we were to that first possibility until I listened to a recent episode of The New York Times podcast Hard Fork, hosted by Times tech columnist Kevin Roose and Casey Newton of Platformer.

“I will make a confession here on this podcast that I have tried to write parts of my column using AI,” Roose said. “I’ve said, ‘I’m sort of stuck on this paragraph. I wonder if it could help me figure out a way to complete this thought.’”

Roose hasn’t been entirely impressed with the results. (He used an app called Lex that he described as a “Google Doc with GPT-3 built in.”)

“Sometimes what it comes up with is passable, but it’s not good,” Roose said. “It’s not something that I would be happy to pass off as my own, even if it were ethical to do so — which I don’t think it would be.”

Our own Joshua Benton came to a similar conclusion after experimenting with GPT-2 back in 2019. Since then, the Microsoft-backed tech company OpenAI has trained its language processing AI on a much larger dataset and introduced a chatbot interface that will bring the technology to many more users than earlier iterations. OpenAI is also developing a watermark that’ll help detect text generated in ChatGPT.

Even with the improvements, Roose said he hasn’t been tempted to include AI-generated writing in his Times column just yet.

“I wouldn’t actually be copying and pasting any of the text verbatim, because it just, frankly, isn’t that unique or interesting or stylish,” Roose said in the episode.

“Maybe it’ll get to a point with GPT-4 where it’s better than I am, and then I’ll have to have some hard thoughts about what I can ethically and spiritually stand outsourcing to the AI,” he added.

Roose envisions using AI help to outline and research his columns. In an earlier Hard Fork episode, the hosts discussed using the tech to generate story ideas, submit broken code for corrections, and create multiple explanations for complicated concepts at different levels of difficulty.

Roose also mentioned another way that AI may help him write his columns.

“One thing that I do when I’m writing is I try to anticipate what people might object to, what good points people might make in response to some argument that I’m making,” Roose said. “I feel like I’m O.K. at that, but a GPT-3 or GPT-4 might be better at it. I might be able to paste in my column and say, ‘What are three counterarguments to this?’”

“Right,” Newton quipped. “Until now, if you wanted to find out why your argument was stupid, you had to tweet out a link to your story.”

Their most recent episode also included Hard Fork’s own predictions. (Newton said “the media’s divorce from Twitter will begin in earnest” in 2023 and Roose claimed to be “medium-confident” that TikTok would be banned in the United States before the year was through.) You can listen or read a transcript here.

This image was generated from the text prompt ‘Teddy bears working on new AI research on the moon in the 1980s.”

Seemingly bound by only your imagination, this latest trend in synthetic media has delighted many, inspired others, and struck fear in some.

new companies will appear that use AI to aggregate and summarize journalism

Google, research firm OpenAI, and AI vendor Stability AI have each developed a text-to-image image generator powerful enough that some observers are questioning whether in the future people will be able to trust the photographic record.

As a computer scientist who specializes in image forensics, I’ve been thinking a lot about this technology: what it’s capable of, how each of the tools have been rolled out to the public, and what lessons can be learned as this technology continues its ballistic trajectory.

Adversarial approach

Although their digital precursor dates back to 1997, the first synthetic images splashed onto the scene just five years ago. In their original incarnation, so-called generative adversarial networks (GANs) were the most common technique for synthesizing images of people, cats, landscapes, and anything else.

A GAN consists of two main parts: generator and discriminator. Each is a type of large neural network, which is a set of interconnected processors roughly analogous to neurons.

Tasked with synthesizing an image of a person, the generator starts with a random assortment of pixels and passes this image to the discriminator, which determines if it can distinguish the generated image from real faces. If it can, the discriminator provides feedback to the generator, which modifies some pixels and tries again. These two systems are pitted against each other in an adversarial loop. Eventually the discriminator is incapable of distinguishing the generated image from real images.

Text-to-image

Just as people were starting to grapple with the consequences of GAN-generated deepfakes — including videos that show someone doing or saying something they didn’t — a new player emerged on the scene: text-to-image deepfakes.

In this latest incarnation, a model is trained on a massive set of images, each captioned with a short text description. The model progressively corrupts each image until only visual noise remains, and then trains a neural network to reverse this corruption. Repeating this process hundreds of millions of times, the model learns how to convert pure noise into a coherent image from any caption.

This photolike image was generated using Stable Diffusion with the prompt “cat wearing VR goggles.”

While GANs are only capable of creating an image of a general category, text-to-image synthesis engines are more powerful. They are capable of creating nearly any image, including images that include an interplay between people and objects with specific and complex interactions, for instance “The president of the United States burning classified documents while sitting around a bonfire on the beach during sunset.”

OpenAI’s text-to-image image generator, DALL-E, took the internet by storm when it was unveiled on January 5, 2021. A beta version of the tool was made available to 1 million users on July 20, 2022. Users around the world have found seemingly endless ways to prompt DALL-E, yielding delightful, bizarre and fantastical imagery.

A wide range of people, from computer scientists to legal scholars and regulators, however, have pondered the potential misuses of the technology. Deep fakes have already been used to create nonconsensual pornography, commit small- and large-scale fraud, and fuel disinformation campaigns. These even more powerful image generators could add jet fuel to these misuses.

Three image generators, three different approaches

Aware of the potential abuses, Google declined to release its text-to-image technology. OpenAI took a more open, and yet still cautious, approach when it initially released its technology to only a few thousand users. It also placed guardrails on allowable text prompts, including no nudity, hate, violence, or identifiable persons. Over time, OpenAI has expanded access, lowered some guardrail, and added more features, including the ability to semantically modify and edit real photographs.

Stability AI took yet a different approach, opting for a full release of their Stable Diffusion with no guardrails on what can be synthesized. In response to concerns of potential abuse, the company’s founder, Emad Mostaque, said, “Ultimately, it’s people’s responsibility as to whether they are ethical, moral and legal in how they operate this technology.”

Nevertheless, the second version of Stable Diffusion removed the ability to render images of NSFW content and children because some users had created child abuse images. In responding to calls of censorship, Mostaque pointed out that because Stable Diffusion is open source, users are free to add these features back at their discretion.

The genie is out of the bottle

Regardless of what you think of Google’s or OpenAI’s approach, Stability AI made their decisions largely irrelevant. Shortly after Stability AI’s open-source announcement, OpenAI lowered its guardrails on generating images of recognizable people. When it comes to this type of shared technology, society is at the mercy of the lowest common denominator — in this case, Stability AI.

Stability AI boasts that its open approach wrestles powerful AI technology away from the few, placing it in the hands of the many. I suspect that few would be so quick to celebrate an infectious disease researcher publishing the formula for a deadly airborne virus created from kitchen ingredients, while arguing that this information should be widely available. Image synthesis does not, of course, pose the same direct threat, but the continued erosion of trust has serious consequences ranging from people’s confidence in election outcomes to how society responds to a global pandemic and climate change.

Moving forward, I believe that technologists will need to consider both the upsides and downsides of their technologies and build mitigation strategies before predictable harms occur. I and other researchers will have to continue to develop forensic techniques to distinguish real images from fakes. Regulators are going to have to start taking more seriously how these technologies are being weaponized against individuals, societies and democracies.

And everyone is going to have to learn how to become more discerning and critical about how they consume information online.

This article has been updated to correct the name of the company Stability AI, which was misidentified.

On October 3, renowned South Korean illustrator Kim Jung Gi passed away unexpectedly at the age of 47. He was beloved for his innovative ink-and-brushwork style of manhwa, or Korean comic-book art, and famous for captivating audiences by live-drawing huge, intricate scenes from memory.

Just days afterward, a former French game developer, known online as 5you, fed Jung Gi’s work into an AI model. He shared the model on Twitter as an homage to the artist, allowing any user to create Jung Gi-style art with a simple text prompt. The artworks showed dystopian battlefields and bustling food markets — eerily accurate in style, and, apart from some telltale warping, as detailed as Jung Gi’s own creations.

The response was pure disdain. “Kim Jung Gi left us less than [a week ago] and AI bros are already ‘replicating’ his style and demanding credit. Vultures and spineless, untalented losers,” read one viral post from the comic-book writer Dave Scheidt on Twitter. “Artists are not just a ‘style.’ They’re not a product. They’re a breathing, experiencing person,” read another from cartoonist Kori Michele Handwerker.

Far from a tribute, many saw the AI generator as a theft of Jung Gi’s body of work. 5you told me that he has received death threats from Jung Gi loyalists and illustrators, and asked to be referred to by his online pseudonym for safety.

Generative AI might have been dubbed Silicon Valley’s “new craze,” but beyond the Valley, hostility and skepticism are already ramping up among an unexpected user base: anime and manga artists. In recent weeks, a series of controversies over AI-generated art — mainly in Japan, but also in South Korea — have prompted industry figures and fans to denounce the technology, along with the artists that use it.

While there’s a long-established culture of creating fan art from copyrighted manga and anime, many are drawing a line in the sand where AI creates a similar artwork. I spoke to generative AI companies, artists, and legal experts, who saw this backlash as being rooted in the intense loyalty of anime and manga circles — and, in Japan, the lenient laws on copyright and data-scraping. The rise of these models isn’t just blurring lines around ownership and liability, but already stoking panic that artists will lose their livelihoods.

“I think they fear that they’re training for something they won’t ever be able to live off because they’re going to be replaced by AI,” 5you told me.

One of the catalysts is Stable Diffusion, a competitor to the AI art model Dall-E, which hit the market on August 22. Stability AI is open-source, which means that, unlike Dall-E, engineers can train the model on any image data set to churn out almost any style of art they desire — no beta invite or subscription needed. 5you, for instance, pulled Jung Gi’s illustrations from Google Images without permission from the artist or publishers, then fed them into Stable Diffusion’s service.

In mid-October, Stability AI, the company behind Stable Diffusion, raised $101 million at a valuation of $1 billion. Looking for a cut of this market, AI startups are building off Stable Diffusion’s open-source code to launch more specialized and refined generators, including several primed for anime and manga art.

Japanese AI startup Radius5 was one of the first companies to touch a nerve when, in August, it launched an art-generation beta called Mimic that targeted anime-style creators. Artists could upload their own work and customize the AI to produce images in their own illustration style; the company recruited five anime artists as test cases for the pilot.

Almost immediately, on Mimic’s launch day, Radius5 released a statement that the artists were being targeted for abuse on social media. “Please refrain from criticizing or slandering creators,” the company’s CEO, Daisuke Urushihara, implored the swarm of Twitter critics. Illustrators decried the service, saying Mimic would cheapen the art form and be used to recreate artists’ work without their permission.

And they were partly right. Just hours after the statement, Radius5 froze the beta indefinitely because users were uploading other artists’ work. Even though this violated Mimic’s terms of service, no restrictions had been built to prevent it. The phrase “AI学習禁止” (“No AI Learning”) lit up Japanese Twitter.

A similar storm gathered around storytelling AI company NovelAI, which launched an image generator on October 3; Twitter rumors rapidly circulated that it was simply ripping human-drawn illustrations from the internet. Virginia Hilton, NovelAI’s community manager, told me that she thought the outrage had to do with how accurately the AI could imitate anime styles.

“I do think that a lot of Japanese people would consider [anime] art a kind of export,” she said. “Finding the capabilities of the [NovelAI] model, and the improvement over Stable Diffusion and Dall-E — it can be scary.” The company also had to pause the service for emergency maintenance. Its infrastructure buckled from a spike in traffic, largely from Japan and South Korea, and a hacking incident. The team published a blog post in English and Japanese to explain how it all works, while scrambling to hire friends to translate their Twitter and Discord posts.

The ripple effect goes on. A Japanese artist was obliged to tweet screenshots showing layers of her illustration software to counter accusations that she was secretly using AI. Two of the country’s most famous VTuber bands requested that millions of social media followers stop using AI in their fan art, citing copyright concerns if their official accounts republished the work. Pixiv, the Japanese online artists’ community, has announced it will be launching tags to filter out AI-generated work in its search feature and in its popularity rankings.

In effect, manga and anime are acting as an early testing ground for AI art-related ethics and copyright liability. The industry has long permitted the reproduction of copyrighted characters through doujinshi (fan-made publications), partly to stoke popularity of the original publications. Even the late Prime Minister Shinzo Abe once weighed in on the unlicensed industry, arguing it should be protected from litigation as a form of parody.

Outside of doujinshi, Japanese law is ordinarily harsh on copyright violations. Even a user who simply retweets or reposts an image that violates copyright can be subject to legal prosecution. But with art generated by AI, legal issues only arise if the output is exactly the same, or very close to, the images on which the model is trained.

“If the images generated are identical…then publishing [those images] may infringe on copyright,” Taichi Kakinuma, an AI-focused partner at the law firm Storia and a member of the economy ministry’s committee on contract guidelines for AI and data, told me. That’s a risk with Mimic and similar generators built to imitate one artist. “Such [a result] could be generated if it is trained only with images of a particular author,” Kakinuma said.

But successful legal cases against AI firms are unlikely, said Kazuyasu Shiraishi, a partner at the Tokyo-headquartered law firm TMI Associates. In 2018, the National Diet, Japan’s legislative body, amended the national copyright law to allow machine-learning models to scrape copyrighted data from the internet without permission, which offers up a liability shield for services like NovelAI.

Whether images are sold for profit or not is largely irrelevant to copyright infringement cases in the Japanese courts, said Shiraishi. But to many working artists, it’s a real fear.

Haruka Fukui, a Tokyo-based artist who creates queer romance anime and manga, admits that AI technology is on track to transform the industry for illustrators like herself, despite recent protests. “There is a concern that the demand for illustrations will decrease and requests will disappear,” she said. “Technological advances have both the benefits of cost reduction and the fear of fewer jobs.”

Fukui has considered using AI herself as an assistive tool, but showed unease when asked if she would give her blessing to AI art generated using her work.

“I don’t intend to consider legal action for personal use,” she said. “[But] I would consider legal action if I made my opinion known on the matter, and if money is generated,” she added. “If the artist rejects it, it should stop being used.”

But the case of Kim Jung Gi shows artists may not be around to give their blessing. “You can’t express your intentions after death,” Fukui admits. “But if only you could ask for the thoughts of the family.”

By its very nature, hype makes it difficult to distinguish illusion from reality. For a reality check, we therefore look to the judgment of the markets themselves, taking measure of products, market sizes, and profits. The resulting numbers suggest that the current tumble in stock prices, tech stock prices in particular, is more than a temporary correction. It is a symptom of a deep disconnect: Tech hype has been distorting people’s behaviors and distracting us from one of the most fundamental economic problems of our time.

Thinking clearly about technological progress versus technological hype requires us to consider the question of why people buy and adopt new technologies in general. A type of academic analysis called the technology acceptance model identifies two notable factors: perceived ease of use and perceived usefulness. That is, we embrace new technologies when they seem easy enough to use and when we believe they will help us do something worthwhile.

As economist Robert Gordon documented in The Rise and Fall of American Growth, Americans adopted a wide variety of technologies that we now take for granted during the “special century” of economic growth between 1870 and 1970. This period featured the advent of a wide variety of pivotal technologies, including steel, running water, machine tools, assembly lines, concrete structures, electric lights and appliances, automobiles, airplanes, pharmaceuticals, computers…the list goes on and on. Many of these technologies enabled individuals and organizations to get more work done with less effort, increasing productivity — the ratio between input and output. As a result, prices for manufactured goods and services fell while output soared. Per capita income in the United States increased by a factor of six between 1870 and 1973, in large part because of these changes.

For reasons still not fully understood, productivity growth came crashing to a halt in the economic difficulties of the 1970s, and it remained low through the 1980s and early 1990s. Then, from 1994 to 2004, productivity briefly rose again, probably in response to the newly commercialized internet, personal computers, and enterprise software that allowed more precise management of business and manufacturing. For a while the dot-com hype seemed justified, but since 2004, productivity growth has been low again, just like in the 70s and 80s.

Much of today’s tech hype is designed to create the impression that the good times of productivity growth are back, or that they never really went away. Innumerable stories tout the looming transformative impact of artificial intelligence (AI), drones, self-driving cars, and whatnot. From our perspective, these technologies aren’t being adopted in transformative ways, largely because they are not allowing us to get that much more done than we did before. Still, measuring productivity is difficult, and people who buy into technology hype can argue that our methods for quantifying productivity are missing important changes. For example, some economic measures, like GDP, do not account for apps that we use for free, like YouTube, Instagram, Facebook, Google Maps, or Waze.

We believe that looking at markets themselves allows us to see through the confusion. Regardless of whether new technologies are truly boosting productivity, if corporations or individuals perceived them to be useful, they would buy them. In reality, we find that people aren’t buying these technologies in large numbers.

One way to cut through the hype is to look at market sizes, a measure of total revenue generated by sales in a given industry, and compare the new technologies to successful ones that came before. For simplicity, let’s look back to the digital technologies from the 1990s dot-com boom. By 2000, revenue generated by e-commerce, internet hardware, and internet software had reached $446 billion, $315 billion, and $282 billion respectively (in 2020 dollars).

Today’s much-hyped technologies do not compare favorably at all: video streaming ($70 billion), big data/algorithms ($46 billion, including companies like Salesforce), smart homes ($20 billion, United States only, including companies like Nest), artificial intelligence ($17 billion), virtual reality ($16 billion), augmented reality ($11 billion in 2019), commercial drones ($6 billion in 2018), and blockchain ($1.9 billion in 2020). The most valuable of these, video streaming, which includes services like pornography, Netflix binges, and cat videos, is highly unlikely to lead to productivity growth; indeed, often enough it distracts us from our work.

When we take the financial measure of individual new technology firms, things look even worse. An ongoing analysis of start-ups by University of Florida economist Jay Ritter has shown that the percentage of startups that were unprofitable during the year prior to their IPO (initial public offering of stock shares) has increased from about 20 percent in the early 1980s to more than 80 percent in the last few years. New firms are simply much less profitable than they used to be. Our evaluation finds that more than 90 percent of today’s big start-ups (those valued at $1 billion or more before they went public) have run cumulative losses over their existence. Uber, the former tech company of the year, has run losses of $29.5 billion. Most of these companies may never climb out of the holes they’ve dug.

A major problem with tech hype (and the journalists who enable it) is that it encourages false optimism. Nobel laureate Robert Shiller describes this “irrational exuberance” in terms of narratives. Investors don’t just look at cold, hard facts such as market sizes and profits; they also follow stories that emphasize intense technological change and big benefits from those changes. One of the most exuberant narratives created the current startup and tech bubble and sustained it even as the promised changes and benefits failed to materialize.

The deeper, even more essential problem with tech hype is that it obscures serious underlying economic issues. Elected officials, civil servants, university professors, and citizens alike have been seduced by technologies that promised sweeping social benefits and economic growth. In the meantime, leaders have neglected fundamentals — low-quality jobs, income stagnation, and inadequate housing — that are the real causes of economic suffering.

According to United Way’s ALICE program, about 40% of working households in the United States now struggle to make ends meet. One big reason: Since the 1970s, and especially since the 1990s, new technologies have not led to the creation of major job-producing industries, despite waves of hype about AI, genetic engineering, nanotechnology, and robotics. If any of these technologies had birthed new industries the way boosters said they would, our economy would be in different shape. Tech hype can be seen as a way to distract us from these failures.

These tough economic realities affect both urban and rural locations and all races and ethnicities, but some more than others. Harvard sociologist William Julius Wilson examined the impact of joblessness on urban Black populations in his 1996 book, When Work Disappears. Too many jobs are low-paying, low-skill positions that do not enable families to thrive; even today, when wages are rising in a tight labor market, incomes are still falling behind inflation. Princeton economists Anne Case and Angus Deaton have demonstrated that whites without college educations in the United States are increasingly dying “deaths of despair,” including suicide, alcoholism, and drug overdoses.

The worst may be yet to come. During the dot-com bust of 2000–2002, many firms went bankrupt and markets hemorrhaged value, but we also ended up with important companies like Amazon. E-commerce was rapidly becoming a part of daily life — even for relatively poor people, even when the dot-com bubble was bursting. When the hype clears and our current, unproductive bubble bursts, we will probably be left only with less.

For example, the urban lifestyle has been propped up by subsidized, unsustainable services, like ride sharing and food and grocery delivery, that are now becoming more expensive even as old services, such as taxis and urban supermarkets, have partly disappeared. Uber burned through investor money like mad, in large part by keeping ride prices artificially low. To become profitable, the company is now increasing its rates and will have to keep doing so until what is left is a pricey service used primarily by the well-to-do. The same applies to the other “sharing economy” app-based companies. What will happen when their services become unaffordable for many people, or when many of these companies disappear entirely? Bankruptcy is a growing possibility as share prices plummet.

In the face of these hard economic realities, the tech-hype machine has latched on to the saddest technologies yet: nonfungible tokens (NFTs), blockchain-based Web 3.0, and Facebook’s “metaverse.” We authors had long wondered what would come along after self-driving cars, AI, and such had lost their luster. We never imagined the answer would be so ridiculous. The poor need better housing, health care, education, and transportation, not an NFT of a goofy cartoon monkey.

Looking at market measures offers a way to see the reality behind the tech hype, but economic data alone cannot show how we can do better. We have a few suggestions.

First and most important, influential figures, including political and journalistic leaders, need to pull back from the dramatic claims of interested parties and examine the larger technological picture. That means talking honestly about which industries are actually improving productivity and creating stable, high-wage jobs. President Obama, who like many Democratic politicians is pretty friendly with Silicon Valley, allowed himself to be taken in by the hype, publicly worrying that AI would soon render many jobs obsolete. If he had asked advisers for realistic guidance on how AI was affecting the economy, he would have seen quite a different picture. He should have focused more on realistic solutions to the deep social problems he cared about — such as poor housing, lack of transportation, and a warming climate — and on ways that these problems can be addressed through technological change.

Academics need to do some strategic rethinking of their own. Universities have issued wild claims about the impact of AI and robots on jobs, often using quantitative methods that are divorced from economic reality. Courses on AI, blockchain, and other new technologies have proliferated, hinting at a conflict of interest between issuing academic projections and earning academic income from courses. Yes, universities have to respond to the market, but they also have the power to influence the market.

Investors and business leaders need to take their social responsibility more seriously. As part of their tech hype, today’s capital pitches and press releases often express a tone of do-goodism, promoting themselves as “agents of change.” It would be better for them to focus on fundamentals like changing market sizes, productivity growth, and job creation, which are what will get us out of trouble if anything will.

And every reader can ask a simple question to avoid being taken in by tech hype: How could this new technology make a positive impact on people’s lives? Consider a boring yet important historical example, machine tools. Emerging in the late 1800s, these tools led to cheaper automobiles, bicycles, construction equipment, and farm equipment, which in turn made food cheaper. Even if they didn’t buy machine tools, average Americans could easily see that their lives were improved through the impact of those tools on productivity. How could cryptocurrencies, NFTs, and the metaverse possibly make the lives of nonusers better? Unfortunately, we authors have trouble finding university professors — even business, economic, and engineering professors — who are asking such simple yet important questions.

Getting back to the fundamentals of technologies and economies will require us to escape our current unproductive bubble of tech hype. Our bet is that, sadly, this process will be painful and will really come about only when the bubble implodes.

Researchers have also asked for dynamic “Transparency APIs” to track and compare these and other changes in real time for reporting reasons, but many have yet to receive the kinds of data they need to conduct the most effective research. For a summary of current research approaches, see “New Approaches to Platform Data Research” from NetGain Partnership. The report points out that even as platforms provide total numbers and categories of information removed, they aren’t informative about “the denominator” of the information total, or what kinds of groups information is and isn’t distributed to. Because of this, there is very little structured information about the efficacy of specific interventions compared to each other. This results in researchers scraping details from product blogs, corporate Twitter threads, and technology reporting.

If these interventions are to have a positive societal impact, we need to be able to measure that impact. This might start with common language, but ultimately we’ll need more to be able to compare interventions to each other. This begins with platforms taking responsibility for reporting these effects and taking ownership of the fact that their intervention decisions have societal effects in the first place. Our prior research surfaced widespread questioning and skepticism of platform intervention processes. In light of this, such ownership and public communication is essential to building trust. That is, platforms can’t simply count on tweaking and A/B testing the color scheme and terminology of existing designs to make the deeper social impacts they appear to seek.

Going forward, we need to examine such patterns and ad hoc goals. We also need to align on what other information is needed and ongoing processes for expanding access to relevant metrics about intervention effects. This includes further analysis of how existing transparency reports are used to understand how they might be more valuable for affecting how users come into contact with content online. Platforms should embrace transparency around the specific goals, tactics, and effects of their misinformation interventions, and take responsibility for reporting on their content interventions and the impact those interventions have.

As a next step, the Partnership on AI is putting together a series of interdisciplinary workshops with our Partners with the ultimate goal of assessing which interventions are the most effective in countering and preventing misinformation — and how we define misinformation in the first place. We’re complementing this work with a survey of Americans’ attitudes towards misinformation interventions. In the meantime, our database serves as a resource to be able to directly compare and evolve interventions in order to help us build a healthier information ecosystem, together.

Contribute to our Intervention Database.

Do you have something to add that we didn’t cover here? We know our list is far from comprehensive, and we want your help to make this a valuable and up-to-date resource for the public. Let us know what we’re missing by emailing aimedia@partnershiponai.org or submitting an intervention to this Airtable form and we’ll get to it as soon as we can. Stay tuned for more updates on future versions of this database and related work.

When an astonishingly realistic deepfake of Tom Cruise spread across the internet in March, many people were quite rightly shocked. Its pinpoint realism suggested artificial intelligence had leapt forward several years.

But one important feature was easily missed. By using the social media handle “deeptomcruise,” the creator was transparent about the fact it was a deepfake.

In a future full of media manipulated by artificial intelligence, we will need methods like this to indicate what is real and what has been faked with AI.1

And this won’t just be a question of ethics. The EU may require that users be told when AI has generated or manipulated something they are viewing.

How should we label AI media in ways that people understand? And how might it backfire?

To start making sense of this phenomenon, we’ve conducted a landscape review of how AI media is currently being labeled. It draws on our previous research on labels, and the insights of many experts in this field, focusing on any media that is generated or manipulated by automated technologies, especially through the use of AI.

In this report, we explore emergent tactics and offer a vocabulary for anyone who wants to label the role AI has played in their creation. In part two, we’ll reflect on the questions, tensions, and dilemmas they raise.

Labels: Covering the media

We use “labels” to refer to any verbal or iconographic item superimposed over content to indicate that it has been manipulated or generated by AI.

Promo watermarks. Watermarks are icons or filters that visually cover content to advertise or cite the tool used to create it. They are also often automatically applied by commercial AI manipulation tools, such as impressions.app, as a means of promotion. This offers insight into the fact that the media was edited and a specific tool used to do so. (At right: Example of a promo watermark applied by Impressions.app; source: TikTok.)

While their effectiveness for different users may vary depending on the language and design, these labels have one simple advantage: If someone saves or screenshots an image or video and recirculates it out of its original context, the watermark travels with it. The more it covers the image or video, the harder it is to remove through editing techniques such as cropping or blurring.

Indeed, this technique has been used by stock image services, such as Shutterstock, for years to prevent the use of images without proper crediting or citation. However, editing out a watermark is certainly not impossible, a limitation that may be countered with stronger media authentication infrastructure, as we will discuss.

Filter tags. Filter tags are labels applied by an app to indicate when one of its filters has been used.

Filter tags have become a familiar convention on TikTok, Snapchat and Instagram, and have a specific advantage over other kinds of labels in that they are interactive. If you click on a filter tag, you can use it, giving you a practical understanding of what the filter does. It’s also possible to see other videos that have used the filter. These are distinguished from promotional labels because they are applied consistently and interactively within a particular platform. Here’s an example of a “Neon Shadow” filter tag, from TikTok.

The disadvantage is that they are not hard-coded onto the media, meaning that if you download a filtered video from TikTok, the tag will be lost. Additionally, the names of the filters tend to offer little insight into the nature and extent of the manipulation itself.

Platform warnings. Twitter’s synthetic and manipulated media policy outlines the conditions for applying warning labels on Tweets, using the same space as the platform’s fact check labels. While not covering the video or images, the placement of the warning in the middle of the post makes it visible simultaneously with the media. As with the filter tags, they are easily removed when the video is downloaded or recorded.

Here’s an example of Twitter’s manipulated media warning, applied to a cheapfake (we have not yet seen an example of its use on AI content).

Non-platform warnings. Non-platform warning labels describe how third parties, such as journalists and fact checkers, cover media to add an explanation or context. This may be in the form of debunking a claim, or simply a label to describe provenance or note manipulation. At right, an example of a warning label from the News Literacy Project.

One example is the use of overlays that cover misinformed or manipulated content to prevent their recirculation when featured in news stories. This can be an effective technique in journalism. First Draft will be publishing guidance as part of this series on signposting online content, with recommendations for journalists on how to design overlays to reduce the amplification of misinformation.

Metadata: Describing the media

We use “metadata” to refer to any description that is manually added to information that accompanies the content.

Title and caption. The title and description of a post provide a space to offer detail on the role of AI. In certain online contexts, they can be highly visible ways of alerting audiences. “The Deepfake Queen,” a video by Channel 4 News, alerts audiences to the use of deepfake technology in the title.

However, titles rarely have much space for detailing the role of AI, especially when needing to also describe something about the media’s content. What’s more, titles and descriptions often do not travel with the media, and the visibility of this metadata changes substantially depending on the platform and screen size.

Profile information. Profile information, such as username and bio, can provide an indication or explanation of the synthetic nature of the account’s content. For example, the “deeptomcruise” TikTok account name provides an indication that the account contains deepfake Tom Cruise videos, for those aware of this usage of the term “deep” as terminology for AI-generated media. Here’s the Instagram profile of the virtual influencer @lilmiquela.

However, this can result in a single explanation for every post on that account, when each one may have used slightly different techniques. On most platforms you also have to click through to the profile page to see bio information, and so this may easily be missed. And because these details are added voluntarily by account creators for various reasons, they may be subtle or satirical in a way that is not widely understood.

Byline. A byline can attribute authorship to AI. It is similar to the username on an account, but it is worth distinguishing because a byline carries significant weight in journalism. An example outside the journalism context could include a signature on a painting — fine art’s equivalent of a byline. An example from a controversial Guardian article is below.

The use of bylines to credit AI authorship has caused controversy because of what’s called the “attribution fallacy,” a term coined by AI researcher Hamid Mattioli Ekbia to describe when people uncritically accept that a task was performed autonomously by AI. A byline can fall afoul of this fallacy by suggesting sole authorship by the AI agent, overstating the capability of the technology and concealing the role of human editors.

Interruption: During the media

We use “interruption” to refer to any indicator that occurs during content display to explain the role of AI in manipulating or generating it. This may be during time-based media (e.g., at the opening of a video) or through delaying access to content (e.g., a pop-up notice).

Pre-roll. In AI-manipulated video, a common tactic is to use a pre-roll label: a notice before the video starts that explains the role of AI.

Pre-roll labels have the advantage of alerting viewers to the role of AI (and more generally that the events didn’t actually occur) before someone watches the video, so they understand what they’re looking at from the start.

The problem is that this opening notice can very easily be edited out and the video reshared as if it were real.

Reveal. A reveal is the opposite of a pre-roll: It tells viewers after they’ve watched a video that it has been AI-manipulated.

The reveal can have great dramatic effect, inducing a state of surprise and awe in the viewer. This can help to emphasize a point, such as the believability of the deepfake.

This is a common tactic. But, even if well-intended, it involves a form of temporary deception, aiming to trick the viewer. This has the risk of being recalled later on as real if this is how it was initially understood. It also relies on someone making it to the end of the video (many people just watch the opening on social media). And, similar to the pre-roll, a reveal is easily edited out.

Interstitial. Interstitial labels interrupt media one or more times to notify the audience of something. This is most commonly used in advertising; longer YouTube videos, for example, use interstitial ads, as do many podcasts. Interstitial labels would make it harder for audiences to ignore or forget the message; harder to join too late or too early and miss it; and harder to edit out the alert.

We have only seen a few of these in the wild (see example from “The Late Show with Stephen Colbert”). We have spoken to technology companies that are conducting internal research on interstitial labels, and are encouraging them to share their findings regarding the impact of those labels more publicly.

Annotation: highlighting the media

We use “annotation” to refer to anything that highlights discrete parts of content, such as areas of an image, durations of a video, or sections of a text.

Annotation. Annotation refers to highlighting specific features of media to indicate where AI manipulation has occurred. This might include highlighting parts of a video that have been manipulated, or could involve narrating over and between snippets of synthetic audio.

1) Ears are nightmare factories, just totally butchered
2) Light source different in the two eyes
3) Teeth are bizarre / wrong
4) Nose just off with nostril and tip of the nose
5) Background is really messed up on the left side

— Aric Toler (@AricToler) June 17, 2020

Annotation is embedded in detection technologies and used by journalists as it provides a tool for precise explanation. This works well for those whose primary subject is the manipulation, but less useful for those who want to focus on content.

One risk is that the degree of confidence implied by precise, thin lines — a common technique of many commercial deepfake detectors — may inspire false confidence that something is AI-manipulated. Such a conclusion, in high-stakes environments, could have real-world impact.

Side-by-side. Side-by-side is another emerging tactic in journalism. It places an unmanipulated piece of media next to manipulated media for comparison by the user.

This tactic allows the viewer to discern the manipulation through comparison. It is also less reliant on long textual explanations, and gives an instinctive feel.

This has been used to explain non-AI manipulation of videos, such as the doctored Nancy Pelosi videos of 2019. It could also be deployed in audio, playing an authentic snippet followed by a manipulated version.

Typography. When it comes to writing, typography —  fonts, for example — offer a way to isolate AI-generated elements. We see this in The Pudding’s part-GPT-3-generated essay “Nothing Breaks like an A.I. Heart,” where AI-generated text is highlighted in a sans serif font (the same tactic used in the part-GPT-3-generated book Pharmako-AI).

This method has the advantage of identifying specific components that are AI-generated in a way that will still be visible when screenshotted and shared. It can also be a way to make the AI manipulation interactive: in The Pudding’s essay, you can click a button to re-generate new text from GPT-3.

Speculation: commentary on the media

“Speculation” refers to the space created for guesses, suggestions and debate about the role of AI media when it isn’t authoritatively labeled.

Another way of looking at the question of labeling AI is what happens when you don’t label media at all. Often a lack of labels creates a kind of data deficit, ushering in speculation. Often this occurs in the comments, with viewers offering their guesses as to whether a piece of media has been manipulated with AI. It could also occur within news articles, blogs or social media posts.

We cannot eradicate speculation, but it is important to recognize the role that labels — or their absence — play in the dynamic of speculation that accompanies videos suspected of being AI-manipulated. This is especially relevant in the context of AI, where our understanding of what is and isn’t possible is constantly being challenged and reset. We further explore risks like these in the second part of this series.

Many uses of AI to manipulate media are, of course, not labeled at all, and perhaps both creators and audiences have little interest in knowing the role of AI media. After all, AI plays some kind of role in almost everything we see online.

One of the most common examples is the iPhone’s “portrait mode,” which uses AI to synthesize a shallow depth of focus by blurring what it detects as the background. This can give the impression that the photo was taken by an expensive SLR camera (though increasingly is suggestive of the iPhone’s portrait mode or another blurring filter).

Examples such as these raise an obvious question: When is it necessary to identify content as being AI-manipulated?

A typology of indicators

We have discovered numerous indicators people and organizations are using to explain the role of AI in the production of content (images, videos, audio or text). To summarize, we’ve grouped these methods into five categories:

• Labels: covering the content
• Metadata: describing the content
• Interruption: before, in between or after the content
• Annotation: highlighting the content
• Discussion: speculating about the content

We have also collated and tagged examples in an Airtable database to facilitate closer analysis and comparison. In some cases, there may be machine-readable metadata (such as Exif), but we focus here on labels that are accessible by general audiences.

Our goal with this short report is not to criticize or endorse any labeling tactic, but to discover what practices are currently in use and to explore what they offer and their limitations.

These examples inevitably raise a number of questions, tensions and trade-offs that are worth exploring. In Part 2 of this series, we’ll explore these considerations for AI indicators.

1. For a more in-depth breakdown of technologies, see Data & Society’s cheap fake and deepfake spectrum.

In its global survey on journalism and artificial intelligence, “New Powers, New Responsibilities,” researchers asked 71 news organizations from 32 countries if and how that currently use AI in their newsrooms and how they expect the technology to impact the news media industry. (Since what exactly constitutes AI can be fuzzy, the report defines it as “a collection of ideas, technologies, and techniques that relate to a computer system’s capacity to perform tasks normally requiring human intelligence.”)

Amazing and inspiring! Our smart news assistant Voitto highlighted in great company in just published #JournalismAI report https://t.co/FYJTotL9kU by @CharlieBeckett @PolisLSE @LSEnews and @GoogleNewsInit Big thanks for the team @yleuutiset @Yleisradio! #journalism #ai pic.twitter.com/WLiZtbIeTR

— Jarno M. Koponen (@ilparone) November 18, 2019

Right now, newsrooms mostly use AI in three areas: news gathering, production, and distribution. Of those surveyed, only 37 percent have an active AI strategy. The survey found that while newsrooms were interested in AI for efficiency and competitive purposes, they said they were mostly motivated by the desire to “help the public cope with a world of news overload and misinformation and to connect them in a convenient way to credible content that is relevant, useful and stimulating for their lives.”

“The hope is that journalists will be algorithmically turbo-charged, capable of using their human skills in new and more effective ways,” Polis founding director Charlie Beckett said in the report. “AI could also transform newsrooms from linear production lines into networked information and engagement hubs that give journalists the structures to take the news industry forward into the data-driven age.”

While most respondents said that AI would be beneficial as long as newsrooms stuck to their ethical and editorial policies, they noted that budget cuts as a result of implementing AI could lower the quality of news produced. They were also concerned about algorithmic bias and the role that technology companies will play in journalism going forward.

“AI technologies will not save journalism or kill it off,” Beckett writes. “Journalism faces a host of other challenges such as public apathy and antipathy, competition for attention, and political persecution…Perhaps the biggest message we should take from this report is that we are at another critical historical moment. If we value journalism as a social good, provided by humans for humans, then we have a window of perhaps 2-5 years, when news organisations must get across this technology.”

Here’s a video summary of the report:

And here is a brief response to the report from Johannes Klingebiel of Süddeutsche Zeitung.

A strategy for #AI adoption will vary depending on the nature of the news organisation. However, the #JournalismAI survey shows there are a few key key elements everyone should consider.

Read the report in full at: https://t.co/7sHArSUaPE pic.twitter.com/47qtPtY09U

— Polis@LSE (@PolisLSE) November 18, 2019

From the cited report: "…news organisations have to change (again). They need to adopt some form of AI strategy. They need to change their workflows, systems, and recruitment. Even then, it’s going to be tough." https://t.co/c8ChRn5uu9

— Aaron Bradley (@aaranged) November 18, 2019

One of the wisest things I have read on this subject, from thoughtful @CharlieBeckett at LSE on AI in #journalism: "we need to know what we are talking about when we talk about AI”. https://t.co/cmZqJppVhn

— PeterBale (@PeterBale) November 18, 2019

Half of respondents use AI for news gathering, two-thirds for production and just over half use AI for distribution. Read the findings of this global survey of journalism and artificial intelligence made by ⁦@CharlieBeckett⁩ at ⁦@PolisLSE⁩ https://t.co/fRTqkZ852u

— Alice Antheaume (@alicanth) November 18, 2019

"No, the robots are not going to take over journalism. Yes, the machines might soon be able to do much routine
journalism labour." This is how starts "Journalism AI" report. The first research on how journalism is using #AI in the newsroom. Check it out: https://t.co/ZSqneQth3s

— Beatriz Gálvez (@beaglvz) November 18, 2019

So whether intentionally or not, the creators of the text-generating algorithm GPT-2 played the PR game brilliantly in February when they announced that, well, it just may be too powerful to release to the general public. That generated a wave of global publicity that is, shall we say, uncommon for new text-generating algorithms. (Elon Musk is involved, you’ll be shocked to learn.)

In any event, now, nine months later, the folks at OpenAI have apparently decided that the infopocalypse is not right around the corner and released its secret superweapon GPT-2 into the wild. They say they have “seen no strong evidence of misuse so far” from more limited releases of the technology.

We're releasing the 1.5billion parameter GPT-2 model as part of our staged release publication strategy.
– GPT-2 output detection model: https://t.co/PX3tbOOOTy
– Research from partners on potential malicious uses: https://t.co/om28yMULL5
– More details: https://t.co/d2JzaENiks pic.twitter.com/O3k28rrE5l

— OpenAI (@OpenAI) November 5, 2019

The alleged threat is not, as some journalists have feared, that this machine is going to eventually cover city council meetings and aggregate viral tweets more efficiently than a human reporter could. It’s that the ease of generating semi-convincing output could make it efficient to pump out thousands or millions of individually tailored misinformation, spam emails, or some other text-based form of grift.

I suppose that’s something legitimate to worry about, but my experiences playing around with the GPT-2 has mostly matched this guy’s: Creating readable prose, sensical prose, and believable prose are still pretty different things.

The full GPT-2 model is finally out.

I dunno, guys. pic.twitter.com/k5dP8V2sa5

— Tomer Ullman (@TomerUllman) November 5, 2019

That something sounds/reads like it was generated by GPT-2 is an interesting new kind of an insult.

— Andrej Karpathy (@karpathy) November 6, 2019

To test out its capabilities, I thought I’d feed it the ledes of a few Nieman Lab stories and see what it autogenerated from there. For each of these below, the actual human-written text is in italics; everything after that is “by” GPT-2. (Note that GPT-2, as used on Talk to Transformer, often ends its output mid-sentence.)

Original story: “Meet The Salt Lake Tribune, 501(c)(3): The IRS has granted nonprofit status to a daily newspaper for the first time,” by Christine Schmidt, published Nov. 4.

Verdict: This is in the broad ballpark of news prose style. It’s interesting it chose to contradict the lede so quickly — specifically saying it rejected 501(c)(3) status. The real Fraser Nelson wouldn’t like that quote, I imagine. Call it a B+.

Original story: “Watch your language: ‘Data voids’ on the web have opened a door to manipulators and other disinformation pushers,” by me, published Nov. 1.

Verdict: Choppy short sentences without much flow or sense, despite sticking in the general topic area. I was sorta hoping it would throw in some more Spanish. I’ll give it a C-.

Original story: “The Daily Wire and black salve show that Facebook’s takedown policies have one thing in common: Inconsistency,” by Laura Hazard Owen, published Nov. 1.

Verdict: Pretty realistic at the sentence level, and maybe the self-contradicting “The problem is that the right-wing site has a left-leaning bias” is actually a galaxy-brain commentary on our current politics. A solid B.

Original story: “Why I’m starting a company to build out a new model for local news in the 2020s,” by Ken Doctor, published Oct. 25.

Verdict: Um, no. There’s no mention of Germany in Ken’s piece and I can’t quite figure out what would make the algorithm want to jump straight to Snowden and Merkel. The fractured lines would make this easy to spot. F.

And finally, to go full ouroboros, let’s feed the lede of this story right here into GPT-2.

Original story: This one, silly

Verdict: I’m sure it was right about to pivot back to text-generation algorithms in a couple more grafs. But this is a very reasonable continuation of the lede (other than that first sentence). B.

GPT-2 is not coming to take the jobs of journalists, as some have worried. Paid reporting jobs generally require a certain level of factuality that the algorithm can’t match.

Is it coming for the “jobs” of fake-news writers, those Macedonian teens who until now have had to generate their propaganda (gasp!) by hand? Probably not. Whether your intention is to make money off ad arbitrage or to elect Donald Trump as president of the United States, the key value-add comes in knowing how to exploit a reader’s emotions, biases, preconceptions, and other lizard-brain qualities that can make a lie really hit home. Baiting that hook remains something an algorithm can reliably do. And it’s not as if “lack of realistic writing in grafs 3 through 12” was a real problem limiting most misinformation campaigns.

But I can see some more realistic impacts here. This quality of generated text could allow you to create a website will what appear to be fully fleshed out archives — pages and pages of cogent text going back years — which might make it seem more legitimate than something more obviously thrown together.

GPT-2’s relative mastery of English could give foreign disinformation campaigns a more authentic sounding voice than whatever the B-team at the Internet Research Agency can produce from watching Parks & Rec reruns.

And the key talent of just about any algorithm is scale — the ability to do something in mass quantities that no team of humans could achieve. As Larry Lessig wrote in 2009 (and Philip Bump reminded us of this week), there’s something about a massive data dump that especially encourages the cherry-picking of facts (“facts”) to support one’s own narrative. Here’s Bump:

In October 2009, he wrote an essay for the New Republic called “Against Transparency,” a provocative title for an insightful assessment of what the Internet would yield. Lessig’s argument was that releasing massive amounts of information onto the Internet for anyone to peruse — a big cache of text messages, for example — would allow people to pick out things that reinforced their own biases…

Lessig’s thesis is summarized in two sentences. “The ‘naked transparency movement’…is not going to inspire change,” he wrote. “It will simply push any faith in our political system over the cliff”…

That power was revealed fully in the 2016 election by one of the targets of the Russia probe: WikiLeaks. The group obtained information stolen by Russian hackers from the Democratic National Committee and Hillary Clinton’s campaign chairman, John Podesta…In October, WikiLeaks slowly released emails from Podesta…Each day’s releases spawned the same cycle over and over. Journalists picked through what had come out, with novelty often trumping newsworthiness in what was immediately shared over social media. Activists did the same surveys, seizing on suggestive (if ultimately meaningless) items. They then often pressured the media to cover the stories, and were occasionally successful…

People’s “responses to information are inseparable from their interests, desires, resources, cognitive capacities, and social contexts,” Lessig wrote, quoting from a book called “Full Disclosure.” “Owing to these and other factors, people may ignore information, or misunderstand it, or misuse it.”

If you wanted to create something as massive as a fake cache of hacked emails, GPT-2 would be of legitimate help — at least as a starting point, producing something that could then be fine-tuned by humans.

The key fact of the Internet is that there’s so much of it. Too much of it for anyone to have a coherent view. If democracy requires a shared set of facts — facts traditionally supplied by professional journalists — the ability to flood the zone with alternative facts could take the bot infestation of Twitter and push it out to the broader world.

Seven organizations are getting a combined $750,000 in funding to help answer these questions, the Ethics and Governance of AI Initiative, a joint project of the MIT Media Lab and Harvard’s Berkman Klein Center, announced Tuesday.

These are the projects receiving funding:

Project: Sidekick
Organization: MuckRock Foundation
Award: $150,000

Newsrooms and researchers are winning access to larger and larger document sets, but getting them is just the start. Understanding what is in those PDFs can be just as challenging, requiring hours of sifting and data entry. Sidekick will offer accessible and intuitive crowdsourcing and machine learning tools to help newsrooms and other groups automate turning documents into data, helping quickly analyze tens of thousands of pages while highlighting sections that might go overlooked.

Project: Reporting from the Epicenter of AI
Organization: Seattle Times
Award: $125,000

The Seattle Times will create a one-year reporting series on artificial intelligence and its implications on society. This work will involve producing major enterprise stories that examine the changing nature of work, assumptions about the jobs Americans will have in the future, and the political and public policy issues that are sparked as changes from AI take hold. From driverless vehicles to advanced robotic systems, technological changes promise to disrupt the nature and structure of work just as dramatically over the next decade as the Internet did over the past quarter century. As it has since the Industrial Revolution, the ongoing transformation of working life will ripple out through society in ways that affect income inequality, culture, notions of community and even social stability. The reporting will engage workers whose lives will be affected by AI technologies and amplify their voices, experiences and perspectives.

Project: Community Media Training to Report on AI Systems
Organization: Craig Newmark Graduate School of Journalism at CUNY
Award: $100,000

This project seeks to train niche media organizations on how to cover artificial intelligence, with an emphasis on how the technology will directly impact the people they serve. The Newmark Journalism school will offer workshops to train journalists on topics from how AI systems can shape health, social and financial policy, to analyzing who benefits and who is affected by how algorithms are coded. The program will include a help desk that can guide journalists on technical issues and questions as they report their stories. The trainings will be available to community journalists in New York City and niche journalists from other parts of the U.S. can apply to attend workshops. The Newmark school will also produce a primer on AI, which will be published in multiple languages.

Project: Automating Public Accountability with AI
Organization: Legal Robot
Award: $100,000

Legal Robot will create a free research tool that journalists and the public can use to find and analyze government contracts, so that they may better understand how public entities use the people’s resources. Legal Robot will use public records laws to request a large set of city, county and state contracts, then automate the extraction and analysis of the data with its machine learning tools. The project will then post both a database and create data visualizations for the public to scrutinize. Visualizations will be created in partnership with TLM Works, a web development training program at the San Quentin prison. The project’s goal is to promote government transparency by providing journalists with the tools and data they need to discover links between government and their contractors, and to scrutinize any fraud, waste, or abuse.

Project: Tattle: Promoting Public Awareness and Combating Misinformation on WhatsApp
Organization: Tattle Civic Technologies
Award: $100,000

In India, as in other developing countries, WhatsApp is one of the most widely-used social media platforms. Information including misinformation spreads quickly on the platform. The effects can be far ranging — from changes in people’s health choices to greater social tension in communities and in extreme cases violence against individuals. Tattle aims to support and scale existing fact-checking efforts by creating channels for sourcing content from WhatsApp users; using machine learning to categorize and classify multilingual, multimedia content circulated on chat apps; and distributing fact checked information so that is accessible to mobile-first audience. In the process Tattle aims to enable more transparent research on misinformation in closed networks.

Project: Robustly Detecting DeepFakes
Organization: Rochester Institute of Technology
Award: $100,000

Researchers at the Rochester Institute of Technology will design and evaluate some of the first approaches for robustly and automatically detecting deepfake videos. These detection techniques will combine vision, audio, and language information, including the synthesis of all three for a comprehensive detection that will be much harder to fool. Videos will be labeled with an “integrity score,” signaling to professionals and consumers where media may have been manipulated. A browser extension will color-code the videos for users depending on their score.

Project: Ethics of Algorithms in Latin America
Organization: Chequeado
Award: $75,000

Chequeado will partner with newsrooms around Latin America to produce an in-depth investigative series on the ethical issues surrounding the implementation of artificial intelligence in the region. Additionally, Chequeado will train local journalists how to cover these emerging technologies and produce a guide with recommendations for journalists covering AI and other relevant issues. This work will be shared within the major journalist networks in Latin America.

The Ethics and Governance in AI Initiative is supported by the John S. and James L. Knight Foundation, Omidyar Network, LinkedIn co-founder Reid Hoffman, and the William and Flora Hewlett Foundation.

That means, of course, that plenty of newsrooms don’t have a data journalist around. But even for those that do, “Data journalism takes many forms, and it’s not always clear from the headline that there is potentially useful data within that document or story,” Simon Rogers, Google News Lab’s data editor, wrote this week. “The way that data is presented can vary as well, and though data tables are often the most useful format for data journalists, it isn’t always easy for Google Search to detect and understand tables of data to surface the most relevant results.”

Google, in partnership with ProPublica, announced that it will be including more data in search results. (Disclosure: I’m on a summer fellowship at Nieman Lab, paid for by the Google News Initiative.) “[T]here is no reason why searching for datasets shouldn’t be as easy as searching for recipes, or jobs, or movies,” Google said in an Google AI blog post last year.

Awesome @GoogleNewsInit collaboration facilitating discovery of public datasets by integrating data journalism—like @ProPublica Nonprofit Explorer's 990 tax return database—directly into @Google search results https://t.co/ua7hTQQWYJ pic.twitter.com/lWZyDJCIO4

— Anna Massoglia (@annalecta) August 1, 2018

From the post:

Based on feedback from 30 of the top data journalists in the world, we identified an opportunity to improve how tabular data appears in Google Search and in doing so make it easier for all people to find the data they’re looking for. It works like this: news organizations that publish data in the form of tables can add additional structured data to make the dataset parts of the page easier to identify for use in relevant Search features. One of the participants, ProPublica has been testing the structured data on its interactive databases (for example, on its Nonprofit Explorer).

News organizations that publish data in the form of tables can now add structured data to the existing html of their pages to make datasets easier to identify and surface in Google Search.

Check out our our developer documentation at https://t.co/4O251Pyfea pic.twitter.com/71CpevnM0M

— Google News Initiative (@GoogleNewsInit) July 31, 2018

The feature joins Google’s other data-accessibility projects, including Google’s Public Data Explorer, Google Trends, the Election DataBot. If you’re a newsroom interested in adding it to your site, you can check out more details here.

Robyn Tomlin has led newsrooms in New York, Texas and North Carolina over the past two decades, and amid a torrent of change she’s noticed a constant. The universe of people trying to deceive journalists keeps expanding.

“When the intent goes up, people will find different technologies to try to help support the mission that they have, which in this case is to mislead us,” says Tomlin, McClatchy’s regional editor for the Carolinas. Previously, she was managing editor of The Dallas Morning News and editor of Digital First Media’s Project Thunderdome.

It’s as if journalists like Tomlin, in newsrooms small and large, have been playing a video game without the ability to change its settings. Every year, they level up into a new class of challenges, with more antagonists, more complicated storylines and an adversarial machine that seems to know their next moves.

The accelerating development of artificial intelligence–driven tools that can manipulate or manufacture from scratch convincing videos, ​photos​, and audio files represents the demarcation of a new level of difficulty. At this point, most of us have seen the ​deepfake​s — the ​Jordan Peele/Obama ​and ​Belgian Trump videos​ and ​more​.

Researchers from the University of Washington have developed tools that can generate realistic video of a person based on audio files. Screenshot from video.

The problem for newsrooms is threefold: how to identify sophisticated manipulations, how to educate audiences without ​inducing apathy and deepening mistrust​, and how to keep the growth of this technology from casting doubt on legitimate and truthful stories.

Perhaps the biggest challenge will be managing the multifaceted levels of difficulty, says Mandy Jenkins, the director of news for Storyful, a social intelligence platform that many newsrooms rely on for verification of user-generated content. Jenkins will be a 2018-19 ​John S. Knight Journalism Fellow​ at Stanford University, where she and others will tackle challenges facing the future of journalism.

“As the AI gets better, it’s going to become pretty seamless. So the thing that we’re going to have to study more is the source,” Jenkins says. “Who is sharing this? What incentive do they have? Where did this really come from? Provenance will have to become more important than ever because you won’t be able to rely on things like location. Corroboration will become important, and maybe more difficult, depending on what the video is, but there should be reporting we can do around that. Alibis, figuring out other stories. Who else has seen this? Who else knows about it? Who filmed it?”

An arms race for the truth

There are ​technology efforts​ to create software that can help with verification and provenance. These include ​software that can detect anomalies​ and digital file signatures that record every step from a photo, video or audio clip’s creation to the point of dissemination, documenting all changes made. However, most of these tools are not yet ready or widely available to news organizations.

Hany Farid​, a digital forensics expert at Dartmouth College, notes that the arms race between better manipulation tools and better detection tools hasn’t worked out in favor of the truth so far. ​“Today, the reality is there’s a handful of people in the world who can probably authenticate this type of digital media, but that obviously doesn’t scale,” he says. “You can’t keep calling three people in the world every time you have to authenticate something.” Not long ago, newsrooms like The New York Times would contact Farid to help authenticate material and give him a day to work on it. Now with hyperfast news cycles, journalists want the intricate analysis done in hours.

The Washington Post has confidence in videos and photos it receives from The Associated Press, Reuters and Storyful, relying “heavily on their verification processes — processes that we understand, that we’ve vetted, that we have a clear idea of what they do to authenticate video,“ says senior editor Micah Gelman, the Post’s director of video. Gelman’s team has yet to develop anywhere near the same level of confidence in user-generated content.

Gelman points to ​misrepresenting images​ and video as the more common form of attempted deception currently, such as photos presented as present-day images of a place like Syria when they are actually from older or unrelated events. Washington Post teams are training with experts to get a handle on how to move beyond verification practices that have typically relied on spotting human-introduced errors. Such errors are things like ​imprecise removal​ of someone or something, ​inaccurate shadows​, missing reflections or breaking the laws of physics in some way, or rough cuts that can be spotted by frame or waveform analysis.

“It’s going to be a while before we really have an understanding of how we work to combat it beyond the traditional methods that we have used for a few years now,” Gelman says. “I think it is a dramatic game changer. We have never as a society faced…the power of this fraud. When you go back to looking at all the people who are still falling for really basic fake articles and things like that, when you add to it now the power of it being something you can watch, with a voice and a face that is of a public figure saying something? Oh, I think this has massive potential to create all kinds of disruptions all over the world.”

In the near future, photo apps that ​perfect images​ automatically or allow alterations like ​changing the time of day​ will be as easy to use as an Instagram filter. Although the manipulation technology is available now, Storyful’s ​Jenkins believes there is still enough time before it becomes widespread for newsrooms to invest in training and public education to offset the impact of high-quality fakes.

A system created by researchers from Google and MIT can automatically retouch images in the style of a professional photographer. It can run on a mobile phone and it’s so fast that it can display retouched images in real-time, so that the photographer can see the final version of the image while still framing. (Courtesy of MIT News)

“It’s not going to be available to everyone at that high-level quality for a while, and hopefully we can sort of stem the tide of public education about it and media education about it,” Jenkins says. “It is up to us to make sure that we as journalists are doing our part to continue to educate and expose and be transparent about that process, since we certainly can’t consistently trust the technology companies. It would be nice to, but I don’t.”

In whatever time they do have to prepare, newsrooms will need to get much better at adopting verification and provenance technology to their workflows and deadline pressures — and figure out how to pay for it. Few newsrooms beyond those with substantial technical resources could even consider developing their own technology solutions. McClatchy’s Tomlin says, “I think it’s a legitimate concern, particularly for small and mid-sized newsrooms that don’t have the experience, the technology, the know-how to be able to do the kind of vetting that needs to get done, or, for that matter, that don’t even necessarily know what to be looking for.”

Most organizations are quite behind at using advanced digital tools. An International Center for Journalists ​survey on the state of technology in global newsrooms​ found that 71 percent of journalists use social media to find stories, but just 11 percent use social media verification tools. Additionally, only five percent of newsroom staff members have degrees in technology-related fields. Overall, the study concludes that journalists are not keeping pace with technology.

The problem is here and now

Experts on digital manipulation convey a greater sense of urgency than the journalists do in preparing for this threat.

“I think none of us are prepared for it, honestly, on many levels. This is not just a journalism issue, it’s not a technology issue, it’s also a social media issue,” says ​Farid, who focuses on digital forensics, image analysis and human perception. He works on problems such as developing tools and practices to detect child exploitation or extremist content.

Farid is part of a ​Defense Advanced Research Project Agency​ project called ​Medifor​ that’s working to put forensic techniques into the hands of law enforcement and journalists. He has also publicly spoken about the need for a ​global cyber ethics commission​ to hold accountable companies that facilitate the spread of misinformation.

Aviv Ovadya​, chief technologist at the Center for Social Media Responsibility, breaks the problem down into production, distribution, and consumption. Production is the creation of the software and apps that make it easier to manipulate or generate deceptive material and the accessibility of those tools. Distribution entails the platforms where people are exposed to manipulations. Consumption is the effect these manipulations have on audiences and how it alters their perception of the world.

Currently, Ovadya says, there isn’t any infrastructure for facilitating responsibility on the part of the engineers creating the technology, whether they are working on behalf of tech companies or in academia. “You need people who are paid, that this is their job. Their job is to understand the impact, the unintended consequences. And that doesn’t currently exist. And it really needs to exist.”

At the same time, companies like Facebook, YouTube, and Twitter need to look beyond advertising-driven business models that reward virality, according to Farid. He argues that the platforms need to acknowledge that they are far more than neutral hosts.

“If it was just hosting [content], we could have a reasonable conversation as to how active a platform should be,” Farid says, adding, “But they’re doing more than that. They are actively promoting it.”

Farid points to the refugee crisis of the ​Rohingya​ people in Myanmar as an example of the dangers. Videos and posts that seek to stoke sentiment against the Rohingya people circulate across social platforms. ​”This is real lives at stake here and these companies have to decide what is their responsibility,” he says. “I think until they decide they’re going to do better, no matter how good we get on other fronts, we’re in deep trouble. Because at the end of the day, these are the platforms that are providing all this material for us.”

When examining distribution from the perspective of newsrooms, none have figured out how to keep up with the speed of viral misinformation rocketing around the ecosystem. Clay Lambert is the editorial director of a five-person newsroom at the Half Moon Bay Review in Northern California, where he says he lacks the people and technology to detect and verify at the speed required. ​“Things could be generated out of whole cloth that would look perfectly normal. So, God bless for thinking that journalism’s going to catch all the bad actors,” he says.

On a local level, the more likely threat stems from exaggerations with the intent to distort community perceptions, hide malfeasance, or spread intolerance. ​”The concern isn’t video or images from events where there are multiple witnesses, but the times when it’s something unique or unusual, something that nobody else caught,” Lambert says.

Regardless of how verification evolves, The Washington Post’s Gelman does not see a future in which newsrooms rely less on user generated content to see what’s happening in the world.

“The problem is caused by an opportunity, right? Everybody’s walking around with a video camera now. Everyone. And it’s their phone. So that is a journalistic opportunity to provide coverage that we never had before,” he says. “We can’t say we’re not going to do that anymore. We can’t take that great storytelling opportunity brought to us by technology. What we need to admit is that like all great storytelling opportunities that come from technology, there is a risk, and it’s our job to make sure, to mitigate that risk and not make mistakes.”

Capacity and transparency

For many newsrooms, risk mitigation looks like a comprehensive series of fact-checking steps and layering in transparency.

Versha Sharma​ ​is the managing editor and senior correspondent at NowThis, a distributed news video company that publishes exclusively on social platforms. ​Sharma emphasizes the importance of NowThis’s multi-tiered system. ​”It is of the highest prioritization for us because we are entirely social video,” Sharma says. “We make sure that fact-checking and verification is at the top of every producer’s mind.”

NowThis frequently shows the ​source of footage​ within its videos and again during the closing credits.

“We ask our producers to credit footage no matter where it came from, whether it’s a paid subscription, whether it’s an individual, whether it’s been crowdsourced — whatever it is. So that our audience can always tell,” Sharma says.

Transparency is one of the tactics most likely to help newsrooms counter the now all-too-common claims of fake news. It’s not just that the technology to create sophisticated manipulations or generated media is being used to deceive. It’s that the very existence of the technology makes it easy for critics to claim ​legitimate reporting and media have been faked​.

“The scariest part is that [new technologies will] be an easy tool for, whether it be politicians or pundits or whoever wants to do it, whoever wants to continue sowing distrust. This is a very easy tool for them,” Sharma says.

Across the board, journalists are thinking about how to educate audiences so they understand both the technology and the ethics of digital video, photo and audio production. Newsrooms have their work cut out for them. Recent research from the American Press Institute’s Media Insight Project on ​what Americans and the news media understand about each other​ concludes,​ “We have a public that doesn’t fully understand how journalists work, and journalism that doesn’t make itself understandable to much of the public.”

Many journalists are starting to realize that there is no final fight to win–only more levels of difficulty.

“I think this is going to be a situation where we get burned and several places get burned and burned hard,” Tomlin says, “before we start really trying to clue in to how serious this is.”

Well, I’m not going to say we avoided doom or gloom entirely — but we did get to have a fruitful discussion of some of the more tech-forward ways the industry is changing. In particular:

— Will blockchain meaningfully change the fundamental questions about how we journalism gets funded? (I’m skeptical.)

— Will AI and bots replace reporters? (Maybe on the fringes, but they’re mainly for scale and speed.)

— What is Apple News planning? (Dunno, but I’m hopeful the mobile OS companies can play a more useful role in news than Facebook does.)

It’s a fun conversation, and I hope you’ll give it a listen here.

The robots may still be a few years away, but the conversation is happening now (at least about today’s AI technology in newsrooms). At Al Jazeera’s Future of Media Leaders’ Summit earlier this month, a group of experts in areas from media to machine learning discussed how their organizations frame the ethics behind (and in front of!) artificial intelligence.

Can a #drone drop a robot into a war zone to report? Doesn’t take a meal break- but can’t empathise with a story! Relationship between man and machine- our future co workers #FMLSummit @AlJazeera pic.twitter.com/Bs2NLWcoNp

— Morwen Williams (@morwenw) March 5, 2018

Ethical AI was one of several topics explored during the gathering in Qatar, focused on data security, the cloud, and how artificial intelligence can automate and augment journalism. (“Data has become more valuable than oil,” Mohamed Abuagla told the audience in the same presentation as the drone-reporter concept.)

AI has already been seeded into the media industry, from surfacing trends for story production to moderating comments. Robotic combat correspondents may still be a far-fetched idea. But with machine learning strengthening algorithms day by day and hour by hour, AI innovations are occurring at a breakneck pace. Machines are more efficient than humans, sure. But in a human-centric field like journalism, how are newsrooms putting AI ethics into practice?

Ali Shah, the BBC’s head of emerging technology and strategic direction, explained his approach to the moral code of AI in journalism. Yaser Bishr, Al Jazeera Media Network’s executive director of digital, also shared some of his thinking on the future of AI in journalism. Here are some of the takeaways:

Ali Shah, the BBC

In both his keynote speech and subsequent panel participation, Shah walked the audience through the business and user implications of infusing AI into parts of the BBC’s production processes. He continued returning to the question of individual agency. “Every time we’re making a judgment about when to apply [machine learning]…what we’re really doing is making a judgment about human capacity,” he said. “Was it right for me to automate that process? When I’m talking about augmenting someone’s role, what judgment values am I augmenting?”

Shah illustrated how the BBC has used AI to perfect camera angles and cuts when filming, search for quotes in recorded data more speedily, and make recommendations for further viewing when the credits are rolling on the BBC’s online player. (The BBC and Microsoft have also experimented with a voice interface AI.) But he emphasized how those AI tools are intended to automate, augment, and amplify human journalists’ work, not necessarily replace or supersede them. “Machine learning is not going to be the answer to every single problem that we face,” he said.

The BBC is proud to be one of the world’s most trusted news brands, and Shah pointed to the need for balance between trust in the organization and individual agency. “We’re going to have to strike a balance between the utility and the effectiveness and the role it plays in society and in our business,” he said. “What we need to do is constantly recognize [that] our role should be giving a little bit of control back to our audience members.”

He also spoke about the need to educate both the engineers designing the AI and the “masses” who are the intended consumers of it. “Journalists are doing a fantastic job at covering this topic,” he said, but “our job as practitioners is to…break this down to the audience so they have control about how machine learning and AI are used to impact them.” (The BBC has published explainer videos about the technology in the past.) “We have to remember, as media, we are gatekeepers to people’s understanding of the modern world.”

“The use of AI changes our behavior. Decisions are influenced and social norms change/evolve but we must make sure we are consciously aware of the effects of machine learning.”- Ali Shah of @BBC at #FMLSummit #AI

— Al Jazeera PR (@AlJazeera) March 6, 2018

“It’s not about slowing down innovation but about deciding what’s at stake,” Shah said. “Choosing your pace is really important.”

Yaser Bishr, Al Jazeera Media Network

Bishr, who helped bring AJ+ to life and has since used Facebook to pull followers onto Al Jazeera’s new Jetty podcast network, also emphasized the need to tread carefully.

“The speed of evolution we are going through in AI far exceeds anything we’ve done before,” Bishr said, talking about the advancements made in the technology at large. “We’re all for innovation, but I think the discussion about regulating the policy needs to go at the same pace.”

In conversation with Shah, Rainer Kellerhais of Microsoft, and Ahmed Elmagarmid of the Qatar Computing Research Institute, Bishr reiterated the risks of AI algorithms putting people into boxes and cited Microsoft’s exiled Twitter bot as an example of input and output bias. “The risk is not only during the training of the machine, but also during the execution of the machine,” he said.

Dr Yaser Bishr of @ALJazeera: “In the context of generating news and information, there is some degree of bias because the info is gathered by a human so there is that editorial interference” #FMLSummit #journalism #AI

— Al Jazeera PR (@AlJazeera) March 6, 2018

Elmagarmid countered his concern about speed: “Things are in motion but things are continuous,” he said calmly. “We have time to adapt to it. We have time to harness it. I think if we look back to the Industrial Revolution, look back to the steam engine…people are always perceiving new technology as threatening.

“At the end of the day you will have [not just] newsrooms, but much better and more efficient and smarter newsrooms,” Elmagarmid said.

“AI is not the Industrial Revolution,” Bishr said, adding to his earlier comments: “We’re not really in a hurry in using AI right now.”

There are already elements of this in quite a few newsrooms but this is the first announcement (I’ve seen) of a large news org rearranging itself around AI…. https://t.co/oK7pZbj158

— emily bell (@emilybell) January 10, 2018

Xinhua News Agency president Cai Mingzhao said Xinhua will build a “new kind of newsroom based on information technology and featuring human-machine collaboration.” The agency has also introduced the “Media Brain” platform to integrate cloud computing, the Internet of Things, AI and more into news production, with potential applications “from finding leads, to news gathering, editing, distribution and finally feedback analysis.”

The agency’s announcement was sparse on details, but it’s the latest component of a deep push into AI by China. Last week the country announced plans for a $2.1 billion AI development park to be built in the next five years as part of its drive to become an AI world leader by 2030. Google has also committed to putting roots in China’s AI scene by opening a research center in Beijing, with Bloomberg quoting Google’s leader of the center Fei-Fei Li: “It will be a small team focused on advancing basic AI research in publications, academic conferences and knowledge exchange.” Microsoft also announced plans to create their own R&D lab for AI in Taiwan and hire 200 researchers over the next five years, investing about $34 million.

“We saw lots of interest in AI in China, and the sector is moving so fast in the country,” Chris Nicholson, former Bloomberg news editor and co-founder of AI startup Skymind, told Digiday. “Beijing supports AI, while Baidu, Alibaba and Tencent are all getting into AI. The U.S. still has the best AI talent, but there are many good engineers and AI researchers in China as well.”

Moving aside from the global AI armsrace, the reverberations from China investing in AI-media could echo in journalism worldwide. A report out today from the Reuters Institute’s Digital News Project on media trends for 2018 highlights some of the advances that Chinese AI journalism has made already:

Executive Editor of Quartz, Zach Seward, recently gave a speech in China at a conference organised by tech giant Tencent. This was turned into a news story by a combination of AI based speech to text software, automatic transcription, and an automated newswriting programme called Dreamwriter. Around 2,500 pieces of news on finance, technology, and sports are created by Dreamwriter daily.

Machine-translated English version of the machine-written Chinese article about a human speech in English https://t.co/01uICBfTR0

— Zach Seward (@zseward) November 19, 2017

Video of Dreamwriter hard at work on the side of the stage during another speech pic.twitter.com/vWBgwfqKRz

— Zach Seward (@zseward) November 19, 2017

With technology having saturated Western markets much of the opportunity for growth is shifting to markets like China and India. But Silicon Valley giants like Google and Facebook face restrictions in China in particular, leaving Asian tech firms driving new ideas at a relentless pace. Without a computer-based legacy to worry about, this is a part of the world that is able to fully embrace mobile first technologies. Increasingly we’ll be looking to the East for innovations in technology in 2018…

Frederic Filloux, author of the Monday Note, has been paying close attention to Toutiao, an app that uses artificial intelligence to aggregate content from around 4,000 traditional news providers as well as bloggers and other personal content. Toutiao has around 120m daily active users and an engagement time of 74 minutes per day. Newsfeeds are constantly updated based on what its machines have learnt about reading preferences, time spent on an article, and location. Toutiao claims to have a user figured out within 24 hours. In Korea, Naver is also looking to add AI recommendations to its mobile services. Line is another mobile news aggregator that is popular in Korea, Taiwan, and Japan. News aggregators like Flipboard and Laserlike48 have made little progress in the US and Europe. But that could change as Toutiao, leaning on its $22 billion valuation, is looking to move aggressively into the Western countries this year.

Chinese companies aren’t necessarily expanding internationally for an international advertising base; as Axios’ Sara Fischer points out, they’re interested in targeting Chinese nationals who have moved elsewhere but still use the same technology to stay in touch back home.

There are also some concerns about what the Chinese government could do with AI journalism: Nina Xiang, the co-founder of the artificial intelligence-based China Money Network, wondered about the potential security and privacy issues from Xinhua’s innovations. “The Media Brain…will raise significant concerns over the protection of personal data privacy, or the lack thereof. The tie-up between Alibaba and China’s state news agency — the first of its kind — creates an all-seeing digital eye that can potentially access data collected from countless surveillance cameras across the nation, estimated to total half a billion in the next three years, Internet of Things (IoT) devices, dashboard-mounted car cameras, air pollution monitoring stations and personal wearable devices. Whether people will be able to give permission for their data being used, or even know its being used, is questionable,” she wrote.

“To use a simple analogy, this partnership is as if Amazon, Paypal, CBS, News Corp and Fox were all working with state and city governments in the United States to share both publicly and privately collected data for the purpose of monitoring for potential news events anywhere, anytime in real time across America.”

While some American organizations are slowly introducing AI to their newsrooms, China’s Xinhua is going all in.

There’s never a shortage of fascinating scholarship in the digital news/social media space. This year, we’re spotlighting 10 of the most compelling academic articles and reports published in 2017, which delve into meaty topics such as venture-backed startups, artificial intelligence, personal branding, and the spread of disinformation. We conferred with a small group of scholars to pick the ones we think you’ll want to know about — and remember, this is just a sample. A big thank you to everybody who contributed suggestions on Twitter.

“Paying for Online News: A comparative analysis of six countries”: From the University of Oxford, published in Digital Journalism. By Richard Fletcher and Rasmus Kleis Nielsen.

This study offers both good news and bad news for publishers struggling to figure out pay models. The researchers used data collected via surveys in six countries, including the United States, to gauge who’s paying for news and who’s willing to pay in the future. The good news: Of those who are not paying for online news now, younger Americans are more willing to pay in the future, possibly because they often already pay for other forms of digital media. The bad news: No more than 2 percent of people surveyed in any country said they are “very likely” to pay for news in the future.

“News Use Across Social Media Platforms 2017”: From the Pew Research Center. By Elisa Shearer and Jeffrey Gottfried.

Throughout the year, the Pew Research Center releases survey-based reports examining journalism and news organizations. This report offers important insights into the role social media plays in distributing and accessing news. Some key takeaways: Almost 70 percent of U.S. adults reported getting news via social media. Meanwhile, a growing number of older adults, people of color, and adults without bachelor’s degrees said they turn to social media sites for news. Minority adults are much more likely than white adults to get news from social media — 74 percent reported doing so in 2017, up from 64 percent in 2016. Interestingly, only 5 percent of adults who go to Snapchat for news also often get news from newspapers.

“Venture-Backed News Startups and the Field of Journalism: Challenges, changes, and consistencies”: From George Washington University, published in Digital Journalism. By Nikki Usher.

How do venture-backed news startups compare themselves to traditional media outlets? This article examines 18 startups, including BuzzFeed, GeekWire, and Vox, to understand how this burgeoning area of digital media is changing journalism’s landscape. Usher interviewed top executives, founders, and others to learn how and why these companies formed as well as details about their editorial visions, technological visions, and plans for making money. The study also explores the rise of algorithms in predicting user behavior, the creation of scalable products, and new roles for journalists within an organization where reporters and technical staff are equals.

“Technology Firms Shape Political Communication: The Work of Microsoft, Facebook, Twitter, and Google With Campaigns During the 2016 U.S. Presidential Cycle”: From the University of North Carolina at Chapel Hill and The University of Utah, published in Political Communication. By Daniel Kreiss and Shannon C. McGregor.

This article offers a behind-the-scenes look at how Facebook, Google, Microsoft, and Twitter collaborated with political campaigns during the 2016 U.S. election season. The paper focuses on their role at the Democratic National Convention in 2016 and in providing extensive consulting services to candidates, including Donald Trump, over the course of the campaign. The researchers found that these technology firms “are increasingly the locus of political knowledge and expertise” in digital and data campaigning. Meanwhile, representatives from each firm said “the growth of their work in electoral politics was driven by the desire for direct revenues from their services and products, for candidates to give their services and platforms greater public visibility, and to establish relationships with legislators.”

“When Reporters Get Hands-On With Robo-Writing: Professionals consider automated journalism’s capabilities and consequences”: From LMU Munich and the University of Zurich, published in Digital Journalism. By Neil Thurman, Konstantin Dörr, and Jessica Kunert.

Media innovators continue to find new ways to integrate artificial intelligence into the newsroom, moving well past using crime stats and structured data from athletic games to generate news reports. While plenty of journalists have weighed in on the trend, most don’t have direct experience using the technology. For this study, researchers conducted workshops with a small group of journalists to show them how to use software to create data-driven news content. After getting hands-on experience, journalists were asked about the potentials and limitations of the technology.

Unsurprisingly, journalists had lots of criticisms — for example, there was concern that automation would make verifying information less likely. Some journalists did see benefits, including time savings and reductions in human error. For several, though, “the experience of creating news items this way was difficult, irritating, and did not utilize their innate abilities.”

“Artificial Intelligence: Practice and Implications for Journalism”: From the Brown Institute for Media Innovation and the Tow Center for Digital Journalism at Columbia Journalism School. By Mark Hansen, Meritxell Roca-Sales, Jon Keegan, and George King.

What problems do journalists and technologists uncover when they brainstorm about AI in newsrooms? This report summarizes a three-hour, wide-ranging discussion between journalists and technologists who gathered last summer for an event organized by Columbia University’s Tow Center for Digital Journalism and the Brown Institute for Media Innovation. Among the important takeaways: A knowledge and communication gap between the technologists who create AI and journalists who use it could “lead to journalistic malpractice.” News outlets need to provide audiences with clear explanations for how AI is used to research and report stories. Also, there needs to be “a concerted and continued effort to fight hidden bias in AI, often unacknowledged but always present, since tools are programmed by humans.”

“Partisanship, Propaganda, and Disinformation: Online Media and the 2016 U.S. Presidential Election”: From the Berkman Klein Center for Internet & Society at Harvard University. By Rob Faris, Hal Roberts, Bruce Etling, Nikki Bourassa, Ethan Zuckerman, and Yochai Benkler.

In this report, researchers examine the composition and behavior of media on the right and left to explain how Donald Trump and Hillary Clinton received differing coverage. The report covers a lot of ground in 142 pages, chock-full of bar charts, network maps, and other data visualizations. It even includes a case study on coverage of the Clinton Foundation. The researchers found that while mainstream media gave mostly negative coverage to both presidential candidates, Trump clearly dominated coverage and was given the opportunity to shape the election agenda.

According to the report, far-right media “succeeded in pushing the Clinton Foundation to the front of the public agenda precisely at the moment when Clinton would have been anticipated to (and indeed did) receive her biggest bounce in the polls: immediately after the Democratic convention.” Researchers also found that while fake news was a problem, it played a relatively small role in the 2016 presidential election. “Disinformation and propaganda from dedicated partisan sites on both sides of the political divide played a much greater role in the election,” the researchers wrote.

“Identity lost? The personal impact of brand journalism”: From The University of Utah and Temple University, published in Journalism. By Avery E. Holton and Logan Molyneux.

Newsrooms urge journalists to use social media to promote their work, interact with sources, and build their professional brands. How does that affect what journalists do on Twitter and Facebook when they’re off the clock? This study is one of several published in 2017 that look at how social media impacts journalists’ identities. This one is important because it lays the groundwork for the others. The authors interviewed 41 reporters and editors at U.S. newspapers to explore the challenges they face in integrating their personal and professional identities on social media. They found that reporters “feel pressure to stake a claim on their beat, develop a presence as an expert in their profession and area of coverage, and act as a representative of the news organization at all times. This leaves little room for aspects of personal identity such as family, faith, or friendship to be shared online.”

“How the news media activate public expression and influence national agendas”: From Harvard University, Florida State University, and MIT, published in Science. By Gary King, Benjamin Schneer, and Ariel White.

Journalism really does contribute to the democratic process and this study provides quantitative evidence. In an experiment involving 48 mostly small media organizations, researchers demonstrated that reporting on a certain policy topic prompts members of the public to take a stand and express their views on the topic more often than they would have if a news article had not been published. Researchers looked at website pageviews and social media posts to gauge impact. Their experiment, according to the researchers, “increased discussion in each broad policy area by ~62.7% (relative to a day’s volume), accounting for 13,166 additional posts over the treatment week.”

“Digital News Report 2017”: From the Reuters Institute for the Study of Journalism at the University of Oxford. By Nic Newman, Richard Fletcher, Antonis Kalogeropoulos, David A. L. Levy, and Rasmus Kleis Nielsen.

This latest annual report from the Reuters Institute offers a global look at digital news consumption based on a survey of more than 70,000 people in 36 countries, including the United States. There are lots of great insights to glean from this 136-page report, which examines such issues as news avoidance, access, distrust, polarization, and sharing. It may (or may not) be surprising that the U.S. ranked 7th highest in the area of news avoidance behind Greece, Turkey, Poland, Croatia, Chile, and Malaysia. Thirty-eight percent of Americans reported avoiding the news “often” or “sometimes.”

Worldwide, the amount of sharing and commenting on news via social media has fallen or stayed about the same the past two years. The U.S., which saw small increases in both habits, is an exception. Another interesting takeaway: Some countries are much more likely to pay for news. In Norway, 15 percent of people surveyed said they made ongoing payments for digital news in the last year, compared to 8 percent in the U.S., 6 percent in Japan, 4 percent in Canada, 3 percent in the United Kingdom and 2 percent in the Czech Republic.

“Take you out of these apps and they become useless,” said Phil Izzo, the deputy chief news editor at The Wall Street Journal. And so when, in recent months, the Journal began redesigning its mobile app, personalization was one of the most important considerations. The ultimate result, released in an iOS 11 update last month, was My WSJ, a feed that’s the second panel after the homescreen and that uses AI to offer a customized list of stories based on users’ previous reading habits.

“We wanted to see how we can thread general app trends into the Journal’s app,” said Jordan Sudy, the Journal’s iOS product director. “People are used to scrolling feeds; the dwell time in the feed has gotten longer — we see people spending a long time in the feed now, relative to what they used to do. People are also getting more and more accustomed to AI recommendations, and we’re seeing that the core Journal reader is interested in AI-generated content, as long as it doesn’t get in the way of what the newsroom is giving them.”

The AI is a big part of the strategy: The Journal really wanted this to be passive personalization. “We were trying to avoid, as much as possible, some Apple News–type screen where you have to select your topics before you jump in,” said Izzo. “Any time you try to get people to set things up, it’s a barrier.”

“This is not the whole Flipboard model where you have to click through five screens [to get your customized feed],” said Sudy. “We don’t have to ask you anything. We just know, by virtue of you being a Journal reader, what you’d like to read and what you should read. You don’t have to tell us anything.” Journal parent company Dow Jones has for months been undertaking the process of tagging content across all of its brands and using those tags to create links between stories; it’s the technology that powers the “Related Stories” feature on desktop, for instance. But this was the first time that the personalization had been brought into the app.

When a user opens the Journal’s app, the first thing they see is the News feed, which looks the same for everyone and is curated by editorial. The My WSJ feed, meanwhile, is populated entirely through AI and doesn’t include human curation. “We wanted you to know very quickly that [News] is what editors are recommending, and [My WSJ] is recommended based on your habits,” said Sudy.

One of the questions that the team had was whether the addition of the My WSJ feed would cannibalize or enhance the presence of the human-curated News feed. Would readers simply swipe past News to get to the stuff aimed directly at them? Though My WSJ has only been around since November 1, data so far seems to suggest that it’s been an enhancement. “It’s not cannibalizing anything. It’s been completely additive,” said Izzo. “There aren’t fewer people going to other sections. They’re just going to this section in addition. We’re seeing increased pageviews.”

The app is the only place where the Journal offers a customized feed. There’s no “recommended for you” section on desktop, at least not yet, though the team is working on tracking logged-in users across web and app, so that the My WSJ feed in the app will ultimately be able to serve up content based on things users had read on the web. The Journal’s app attracts “the corest of core readers,” Izzo pointed out, who are already interested in going to sections and landing pages, whereas a lot of web traffic comes sideways from search, so maintaining a personalized feed for them would be more difficult. “The app is that playground where we can try things that somebody coming in from social wouldn’t be interested in,” he said. (The New York Times, meanwhile, is taking a different approach, doing more subtle personalization on desktop.)

In addition to the passive personalization offered in My WSJ, the team is looking cautiously at some forms of active personalization, though Sudy remains wary of making users do too much work. Users can now follow individual Journal journalists and receive notifications when they publish new stories; with the relaunch of the paper’s Markets Data Center, the team plans to let users get alerts on news about individual companies, as well.

One big challenge for news companies is how granular they can or should get with push notifications. Are there any plans to send out push alerts to readers when new stories come up that they, specifically, might be interested in?

“We have not yet done automatic push alerts based on reading behavior, and we’d tread very lightly with that,” said Izzo, who is wary of overloading readers. Still, he and Sudy said it wouldn’t be a difficult feature to build. “We could probably make it happen pretty easily. We’ll think about that.”

A report out this week from the Tow Center for Digital Journalism looks at how AI can be adapted to journalism. It summarizes a previously off-the-record meeting held back in June by the Tow Center and the Brown Institute for Media Innovation. (Full disclosure: Nieman Lab director Joshua Benton was part of the meeting.)

Among the report’s findings:

— AI “helps reporters find and tell stories that were previously out of reach or impractical.” Three areas where AI can be particularly helpful in the newsroom: “Finding needles in haystacks” (discovering things in data that humans can’t, which humans can then fact-check); identifying trends or outliers; and as a subject of a story itself: “Because they are built by humans, algorithms harbor human bias — and by examining them, we can discover previously unseen bias.”

— AI can deliver much more personalized news — for good and bad. AI could be used to monitor readers’ likes and dislikes, ultimately shaping stories to people’s individual interests. But, as one participant cautioned:

The first stage of personalization is recommending articles; the long-term impact is filter bubbles. The next step is using NLP [Natural Language Processing] to shape an article to exactly the way you want to read it. Tone, political stance, and many other things. At that point, journalism becomes marketing. We need to be very aware that too much personalization crosses the line into a different activity.

Another concern is that, if articles become too personalized, the public record is at risk: “When everyone sees a different version of a story, there is no authoritative version to cite.”

— AI brings up new ethical considerations. Participants agreed that news organizations need to disclose when AI has been used in creating a story, but “that description must be translated into non-technical terms, and told in a concise manner that lets readers understand how AI was used and how choices were made.” There’s a need for best practices around disclosures.

Also, most AI tools aren’t built specifically with newsrooms (or their editorial values) in mind. One engineer said: “A lot of these questions currently seem impenetrable to us engineers because we don’t understand the editorial values at a deep level, so we can’t model them. Engineers don’t necessarily think of the systems they are building as embodying editorial values, which is an interesting problem. The way a system like this is built does not reflect this underlying goal.”

The full report is here.

This summer, as a Knight Visiting Nieman Fellow at Harvard, I’ve been looking at another technology I think could lead to a similar step change in how publishers relate to their audiences: AI-driven voice interfaces, such as Amazon’s Alexa, Google’s Home and Assistant, Microsoft’s Cortana, and Apple’s upcoming HomePod. The more I’ve spoken to the editorial and technical leads building on these platforms in different news organizations, as well as the tech companies developing them, the more I’ve come to this view: This is potentially bigger than the impact of the iPhone. In fact, I’d describe these smart speakers and the associated AI and machine learning that they’ll interface with as the huge burning platform the news industry doesn’t even know it’s standing on.

This wasn’t how I planned to open this piece even a week before my Nieman fellowship ended. But as I tied together the research I’d done with the conversations I’d had with people across the industry, something became clear: As an industry, we’re far behind the thinking of the technology companies investing heavily in AI and machine learning. Over the past year, the CEOs of Google, Microsoft, Facebook, and other global tech giants have all said, in different ways, that they now run “AI-first” companies. I can’t remember a single senior news exec ever mentioning AI and machine learning at any industry keynote address over the same period.

Of course, that’s not necessarily surprising. “We’re not technology companies” is a refrain I’ve heard a lot. And there are plenty of other important issues to occupy industry minds: the rise of fake news, continued uncertainty in digital advertising, new tech such as VR and AR, and the ongoing conundrum of responding to the latest strategic moves of Facebook.

But as a result of all these issues, AI is largely being missed as an industry priority; to switch analogies, it feels like we’re the frog being slowly boiled alive, not perceiving the danger to itself until it’s too late to jump out.

“In all the speeches and presentations I’ve made, I’ve been shouting about voice AI until I’m blue in the face. I don’t know to what extent any of the leaders in the news industry are listening,” futurist and author Amy Webb told me. As she put it in a piece she wrote for Nieman Reports recently:

Talking to machines, rather than typing on them, isn’t some temporary gimmick. Humans talking to machines — and eventually, machines talking to each other — represents the next major shift in our news information ecosystem. Voice is the next big threat for journalism.

My original goal for this piece was to share what I’d learned — examples of what different newsrooms are trying with smart speakers and where the challenges and opportunities lie. There’s more on all that below. But I first want to emphasize the critical and urgent nature of what the news industry is about to be confronted with, and how — if it’s not careful — it’ll miss the boat just as it did when the Internet first spread from its academic cocoon to the rest of the world. Later, I’ll share how I think the news industry can respond.

Talking to objects isn’t weird any more

In the latest version of her annual digital trends report, Kleiner Perkins’ Mary Meeker revealed that 20 percent of all Google search was now happening through voice rather than typing. Sales of smart speakers like Amazon’s Echo were also increasing fast:

It’s becoming clear that users are finding it useful to interact with devices through voice. “We’re treating voice as the third wave of technology, following the point-and-click of PCs and touch interface of smartphones,” Francesco Marconi, a media strategist at Associated Press, told me. He recently coauthored AP’s report on how artificial intelligence will impact journalism. The report gives some excellent insights into the broader AI landscape, including automation of content creation, data journalism through machine learning, robotic cameras, and media monitoring systems. It highlighted smart speakers as a key gateway into the world of AI.

Since the release of the Echo, a number of outlets have tried to learn what content works (or doesn’t) on this class of devices. Radio broadcasters have been at an understandable advantage, being able to adapt their content relatively seamlessly.

In the U.S., NPR was among the first launch partners on these platforms. Ha-Hoa Hamano, a senior product manager at NPR working on voice AI, described its hourly newscast as “the gateway to NPR’s content.”

“We’re very bullish on the opportunity with voice,” Hamano said. She cited research showing 32 percent of people aged 18 to 34 don’t own a radio in their home — “which is a terrifying stat when you’re trying reach and grow audience. These technologies allow NPR to fit into their daily routine at home — or wherever they choose to listen.”

NPR was available at launch on the Echo and Google Home, and will be soon on Apple’s HomePod. “We think of the newscast as the gateway to the rest of NPR’s news and storytelling,” she said. “It’s a low lift for us to get the content we already produce onto these platforms. The challenge is finding the right content for this new way of listening.”

The API that drives NPR made it easy for Hamano’s team to integrate the network’s content into Amazon’s system. NPR’s skills — the voice-driven apps that Amazon’s voice assistant Alexa recognizes — can respond to requests like “Alexa, ask NPR One to recommend a podcast” or “Alexa, ask NPR One to play Hidden Brain.”

The Washington Post — owned by Amazon CEO Jeff Bezos — is also an early adopter in running a number of experiments on Amazon’s and Google’s smart speaker platforms. Senior product manager Joseph Price has been leading this work. “I think we’re at the early stages of what I’d call ambient computing — technology that reduces the ‘friction’ between what we want and actually getting it in terms of our digital activity,” he said. “It will actually mean we’ll spend less time being distracted by technology, as it effectively recedes into the background as soon as we are finished with it. That’s the starting point for us when we think about what voice experiences will work for users in this space.”

Not being a radio broadcaster, the Post has had to experiment with different forms of audio — from using Amazon’s Alexa automated voices on stories from its website to a Post reporter sharing a particular story in their own voice. Other experiments have included launching an Olympics skill, where users could ask the Post who had won medals during last year’s Olympics. That was an example of something that didn’t work, though — Amazon built the same capability into the main Alexa platform soon afterwards itself.

“That was a really useful lesson for us,” Price said. “We realized that in big public events like these, where there’s an open data set about who has won what, it made much more sense for a user to just ask Alexa who had won the most medals, rather than specifically asking The Washington Post on Alexa the same question.” That’s a broader lesson: “We have to think about what unique or exclusive information, content, or voice experience can The Washington Post specifically offer that the main Alexa interface can’t.”

One area that Price’s team is currently working on is the upcoming release of notifications on both Amazon’s Alexa and Google’s Home platforms. For instance, if there’s breaking news, the Post will be able to make a user’s Echo chime and flash green, at which point the user can ask “Alexa, what did I miss?” or “Alexa, what are my notifications?” Users will have to opt in before getting alerts to their device, and they’ll be able to disable alerts temporarily through a do-not-disturb mode.

Publishers like the Post that produce little or no native audio content have to work out the right way of presenting their text-based content on a voice-driven platform. One option is to allow Alexa to read stories that have been published; that’s easy to scale up. The other is getting journalists to voice articles or columns or create native audio for the platform. That’s much more difficult to scale, but several news organizations told me initial audience feedback suggests this is users’ preferred experience.

For TV broadcasters like CNN, early experiments have focused on trying to figure out when their users would most want to listen to a bulletin — as opposed to watching one — and how much time they might have specifically to do so via a smart speaker. Elizabeth Johnson, a senior editor at CNN Digital, has been leading the work on developing flash-briefing content for these platforms.

“We assumed some users would have their device in the kitchen,” she said. “This led us to ask, what are users probably doing in the kitchen in the morning? Making breakfast. How long does it take to make a bagel? Five minutes. So that’s probably the amount of time a user has to listen to us, so let’s make sure we can update them in less than five minutes. For other times of the day, we tried to understand what users might be doing: Are they doing the dishes? Are they watching a commercial break on TV or brushing their teeth? We know that we’re competing against a multitude of options, so what sets us apart?”

With Amazon’s recent release of the Echo Show — which has a built-in screen — CNN is taking the “bagel time” philosophy to developing a dedicated video news briefing at the same length as its audio equivalent.

CNN is also thinking hard about when notifications will and won’t work. “If you send a notification at noon, but the user doesn’t get home until 6 p.m., does it make sense for them to see that notification?” Johnson asked. “What do we want our users to hear when they come home? What content do we have that makes sense in that space, at that time? We already consider the CNN mobile app and Apple News alerts to be different, as are banners on CNN.com — they each serve different purposes. Now, we have to figure out how to best serve the audience alerts on these voice-activated platforms.”

What’s surprised many news organizations is how broad the age range of their audiences are on smart speakers. Early adopters in this space are very different from early adopters of other technologies. Many didn’t buy these smart speakers themselves, but were given them as gifts, particularly around Christmas. The fact there’s very little learning curve to use them means the technical bar is much lower. Speaking to the device is intuitive.

Edison Research was recently commissioned by NPR to find out more about what these users are doing with these devices. Music was unsurprisingly at the top of the reasons why they use these devices, but coming in second was to “ask questions without needing to type.” Also high up was an interest to listen to news and information — encouraging for news organizations.

While screens aren’t going away — people will always want to see and touch things — there’s no doubt that voice as an interface for devices is already becoming ingrained as a natural behavior among our audiences. If you’re not convinced, watch children interact with smart speakers: Just as we’ve seen the first Internet-connected generation grow up, we’re about to see the “voice generation” arrive feeling completely at ease with this way of engaging with technology.

The NPR–Edison research has also highlighted this trend. Households with kids that have smart speakers say engagement is high with these devices. Unlike phones or tablet, smart speakers are communal experiences — which also raises the likelihood of families spending time together, whether for education or entertainment purposes.

(It’s worth noting here that there have been some concerns raised about whether children asking for — or demanding — content from a device without saying “please” or “thank you” could have downsides. As San Francisco VC and dad Hunter Walk put it last year: “Amazon Echo is magical. It’s also turning my kid into an asshole.” To curb this, skills or apps for children could be designed in the future with voice responses requiring politeness.)

For the BBC, where I work, developing a voice-led digital product for children is an exciting possibility. It already has considerable experience of content for children on TV, radio, online and digital.

“Offering the ability to seamlessly navigate our rich content estate represents a great opportunity for us to forge a closer relationship with our audience and to serve them better,” Ben Rosenberg, senior distribution manager at the BBC, said. “The current use cases for voice suggest there is demand that sits squarely in the content areas where we consistently deliver on our ambitions — radio, news, and children’s genres.”

BBC News recently formed a working group to rapidly develop prototypes for new forms of digital audio using voice as the primary interface. Expect to hear more about this in the near future.

Rosenberg also highlights studies that have found voice AI interfaces appeared to significantly increase consumption of audio content. This is something that came out strongly in the NPR-Edison research too:

Owning a smart speaker can lead to a sizeable increase in consumption of music, news and talk content, podcasts, and audiobooks. Media organizations that have such content have a real opportunity if they can figure out how to make it as easily accessible through these devices as possible. That’s where we get to the tricky part.